1 /* 2 ** 2001 September 15 3 ** 4 ** The author disclaims copyright to this source code. In place of 5 ** a legal notice, here is a blessing: 6 ** 7 ** May you do good and not evil. 8 ** May you find forgiveness for yourself and forgive others. 9 ** May you share freely, never taking more than you give. 10 ** 11 ************************************************************************* 12 ** This module contains C code that generates VDBE code used to process 13 ** the WHERE clause of SQL statements. This module is responsible for 14 ** generating the code that loops through a table looking for applicable 15 ** rows. Indices are selected and used to speed the search when doing 16 ** so is applicable. Because this module is responsible for selecting 17 ** indices, you might also think of this module as the "query optimizer". 18 */ 19 #include "sqliteInt.h" 20 #include "whereInt.h" 21 22 /* Forward declaration of methods */ 23 static int whereLoopResize(sqlite3*, WhereLoop*, int); 24 25 /* Test variable that can be set to enable WHERE tracing */ 26 #if defined(SQLITE_TEST) || defined(SQLITE_DEBUG) 27 /***/ int sqlite3WhereTrace = 0; 28 #endif 29 30 31 /* 32 ** Return the estimated number of output rows from a WHERE clause 33 */ 34 u64 sqlite3WhereOutputRowCount(WhereInfo *pWInfo){ 35 return sqlite3LogEstToInt(pWInfo->nRowOut); 36 } 37 38 /* 39 ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this 40 ** WHERE clause returns outputs for DISTINCT processing. 41 */ 42 int sqlite3WhereIsDistinct(WhereInfo *pWInfo){ 43 return pWInfo->eDistinct; 44 } 45 46 /* 47 ** Return TRUE if the WHERE clause returns rows in ORDER BY order. 48 ** Return FALSE if the output needs to be sorted. 49 */ 50 int sqlite3WhereIsOrdered(WhereInfo *pWInfo){ 51 return pWInfo->nOBSat; 52 } 53 54 /* 55 ** Return the VDBE address or label to jump to in order to continue 56 ** immediately with the next row of a WHERE clause. 57 */ 58 int sqlite3WhereContinueLabel(WhereInfo *pWInfo){ 59 assert( pWInfo->iContinue!=0 ); 60 return pWInfo->iContinue; 61 } 62 63 /* 64 ** Return the VDBE address or label to jump to in order to break 65 ** out of a WHERE loop. 66 */ 67 int sqlite3WhereBreakLabel(WhereInfo *pWInfo){ 68 return pWInfo->iBreak; 69 } 70 71 /* 72 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to 73 ** operate directly on the rowis returned by a WHERE clause. Return 74 ** ONEPASS_SINGLE (1) if the statement can operation directly because only 75 ** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass 76 ** optimization can be used on multiple 77 ** 78 ** If the ONEPASS optimization is used (if this routine returns true) 79 ** then also write the indices of open cursors used by ONEPASS 80 ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data 81 ** table and iaCur[1] gets the cursor used by an auxiliary index. 82 ** Either value may be -1, indicating that cursor is not used. 83 ** Any cursors returned will have been opened for writing. 84 ** 85 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is 86 ** unable to use the ONEPASS optimization. 87 */ 88 int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){ 89 memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2); 90 #ifdef WHERETRACE_ENABLED 91 if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){ 92 sqlite3DebugPrintf("%s cursors: %d %d\n", 93 pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI", 94 aiCur[0], aiCur[1]); 95 } 96 #endif 97 return pWInfo->eOnePass; 98 } 99 100 /* 101 ** Move the content of pSrc into pDest 102 */ 103 static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){ 104 pDest->n = pSrc->n; 105 memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0])); 106 } 107 108 /* 109 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet. 110 ** 111 ** The new entry might overwrite an existing entry, or it might be 112 ** appended, or it might be discarded. Do whatever is the right thing 113 ** so that pSet keeps the N_OR_COST best entries seen so far. 114 */ 115 static int whereOrInsert( 116 WhereOrSet *pSet, /* The WhereOrSet to be updated */ 117 Bitmask prereq, /* Prerequisites of the new entry */ 118 LogEst rRun, /* Run-cost of the new entry */ 119 LogEst nOut /* Number of outputs for the new entry */ 120 ){ 121 u16 i; 122 WhereOrCost *p; 123 for(i=pSet->n, p=pSet->a; i>0; i--, p++){ 124 if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){ 125 goto whereOrInsert_done; 126 } 127 if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){ 128 return 0; 129 } 130 } 131 if( pSet->n<N_OR_COST ){ 132 p = &pSet->a[pSet->n++]; 133 p->nOut = nOut; 134 }else{ 135 p = pSet->a; 136 for(i=1; i<pSet->n; i++){ 137 if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i; 138 } 139 if( p->rRun<=rRun ) return 0; 140 } 141 whereOrInsert_done: 142 p->prereq = prereq; 143 p->rRun = rRun; 144 if( p->nOut>nOut ) p->nOut = nOut; 145 return 1; 146 } 147 148 /* 149 ** Return the bitmask for the given cursor number. Return 0 if 150 ** iCursor is not in the set. 151 */ 152 Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){ 153 int i; 154 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 ); 155 for(i=0; i<pMaskSet->n; i++){ 156 if( pMaskSet->ix[i]==iCursor ){ 157 return MASKBIT(i); 158 } 159 } 160 return 0; 161 } 162 163 /* 164 ** Create a new mask for cursor iCursor. 165 ** 166 ** There is one cursor per table in the FROM clause. The number of 167 ** tables in the FROM clause is limited by a test early in the 168 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] 169 ** array will never overflow. 170 */ 171 static void createMask(WhereMaskSet *pMaskSet, int iCursor){ 172 assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); 173 pMaskSet->ix[pMaskSet->n++] = iCursor; 174 } 175 176 /* 177 ** Advance to the next WhereTerm that matches according to the criteria 178 ** established when the pScan object was initialized by whereScanInit(). 179 ** Return NULL if there are no more matching WhereTerms. 180 */ 181 static WhereTerm *whereScanNext(WhereScan *pScan){ 182 int iCur; /* The cursor on the LHS of the term */ 183 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */ 184 Expr *pX; /* An expression being tested */ 185 WhereClause *pWC; /* Shorthand for pScan->pWC */ 186 WhereTerm *pTerm; /* The term being tested */ 187 int k = pScan->k; /* Where to start scanning */ 188 189 while( pScan->iEquiv<=pScan->nEquiv ){ 190 iCur = pScan->aiCur[pScan->iEquiv-1]; 191 iColumn = pScan->aiColumn[pScan->iEquiv-1]; 192 if( iColumn==XN_EXPR && pScan->pIdxExpr==0 ) return 0; 193 while( (pWC = pScan->pWC)!=0 ){ 194 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){ 195 if( pTerm->leftCursor==iCur 196 && pTerm->u.leftColumn==iColumn 197 && (iColumn!=XN_EXPR 198 || sqlite3ExprCompare(pTerm->pExpr->pLeft,pScan->pIdxExpr,iCur)==0) 199 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin)) 200 ){ 201 if( (pTerm->eOperator & WO_EQUIV)!=0 202 && pScan->nEquiv<ArraySize(pScan->aiCur) 203 && (pX = sqlite3ExprSkipCollate(pTerm->pExpr->pRight))->op==TK_COLUMN 204 ){ 205 int j; 206 for(j=0; j<pScan->nEquiv; j++){ 207 if( pScan->aiCur[j]==pX->iTable 208 && pScan->aiColumn[j]==pX->iColumn ){ 209 break; 210 } 211 } 212 if( j==pScan->nEquiv ){ 213 pScan->aiCur[j] = pX->iTable; 214 pScan->aiColumn[j] = pX->iColumn; 215 pScan->nEquiv++; 216 } 217 } 218 if( (pTerm->eOperator & pScan->opMask)!=0 ){ 219 /* Verify the affinity and collating sequence match */ 220 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){ 221 CollSeq *pColl; 222 Parse *pParse = pWC->pWInfo->pParse; 223 pX = pTerm->pExpr; 224 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){ 225 continue; 226 } 227 assert(pX->pLeft); 228 pColl = sqlite3BinaryCompareCollSeq(pParse, 229 pX->pLeft, pX->pRight); 230 if( pColl==0 ) pColl = pParse->db->pDfltColl; 231 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){ 232 continue; 233 } 234 } 235 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0 236 && (pX = pTerm->pExpr->pRight)->op==TK_COLUMN 237 && pX->iTable==pScan->aiCur[0] 238 && pX->iColumn==pScan->aiColumn[0] 239 ){ 240 testcase( pTerm->eOperator & WO_IS ); 241 continue; 242 } 243 pScan->k = k+1; 244 return pTerm; 245 } 246 } 247 } 248 pScan->pWC = pScan->pWC->pOuter; 249 k = 0; 250 } 251 pScan->pWC = pScan->pOrigWC; 252 k = 0; 253 pScan->iEquiv++; 254 } 255 return 0; 256 } 257 258 /* 259 ** Initialize a WHERE clause scanner object. Return a pointer to the 260 ** first match. Return NULL if there are no matches. 261 ** 262 ** The scanner will be searching the WHERE clause pWC. It will look 263 ** for terms of the form "X <op> <expr>" where X is column iColumn of table 264 ** iCur. The <op> must be one of the operators described by opMask. 265 ** 266 ** If the search is for X and the WHERE clause contains terms of the 267 ** form X=Y then this routine might also return terms of the form 268 ** "Y <op> <expr>". The number of levels of transitivity is limited, 269 ** but is enough to handle most commonly occurring SQL statements. 270 ** 271 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with 272 ** index pIdx. 273 */ 274 static WhereTerm *whereScanInit( 275 WhereScan *pScan, /* The WhereScan object being initialized */ 276 WhereClause *pWC, /* The WHERE clause to be scanned */ 277 int iCur, /* Cursor to scan for */ 278 int iColumn, /* Column to scan for */ 279 u32 opMask, /* Operator(s) to scan for */ 280 Index *pIdx /* Must be compatible with this index */ 281 ){ 282 int j = 0; 283 284 /* memset(pScan, 0, sizeof(*pScan)); */ 285 pScan->pOrigWC = pWC; 286 pScan->pWC = pWC; 287 pScan->pIdxExpr = 0; 288 if( pIdx ){ 289 j = iColumn; 290 iColumn = pIdx->aiColumn[j]; 291 if( iColumn==XN_EXPR ) pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr; 292 } 293 if( pIdx && iColumn>=0 ){ 294 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; 295 pScan->zCollName = pIdx->azColl[j]; 296 }else{ 297 pScan->idxaff = 0; 298 pScan->zCollName = 0; 299 } 300 pScan->opMask = opMask; 301 pScan->k = 0; 302 pScan->aiCur[0] = iCur; 303 pScan->aiColumn[0] = iColumn; 304 pScan->nEquiv = 1; 305 pScan->iEquiv = 1; 306 return whereScanNext(pScan); 307 } 308 309 /* 310 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 311 ** where X is a reference to the iColumn of table iCur and <op> is one of 312 ** the WO_xx operator codes specified by the op parameter. 313 ** Return a pointer to the term. Return 0 if not found. 314 ** 315 ** If pIdx!=0 then search for terms matching the iColumn-th column of pIdx 316 ** rather than the iColumn-th column of table iCur. 317 ** 318 ** The term returned might by Y=<expr> if there is another constraint in 319 ** the WHERE clause that specifies that X=Y. Any such constraints will be 320 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 321 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11 322 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10 323 ** other equivalent values. Hence a search for X will return <expr> if X=A1 324 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>. 325 ** 326 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 327 ** then try for the one with no dependencies on <expr> - in other words where 328 ** <expr> is a constant expression of some kind. Only return entries of 329 ** the form "X <op> Y" where Y is a column in another table if no terms of 330 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 331 ** exist, try to return a term that does not use WO_EQUIV. 332 */ 333 WhereTerm *sqlite3WhereFindTerm( 334 WhereClause *pWC, /* The WHERE clause to be searched */ 335 int iCur, /* Cursor number of LHS */ 336 int iColumn, /* Column number of LHS */ 337 Bitmask notReady, /* RHS must not overlap with this mask */ 338 u32 op, /* Mask of WO_xx values describing operator */ 339 Index *pIdx /* Must be compatible with this index, if not NULL */ 340 ){ 341 WhereTerm *pResult = 0; 342 WhereTerm *p; 343 WhereScan scan; 344 345 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); 346 op &= WO_EQ|WO_IS; 347 while( p ){ 348 if( (p->prereqRight & notReady)==0 ){ 349 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){ 350 testcase( p->eOperator & WO_IS ); 351 return p; 352 } 353 if( pResult==0 ) pResult = p; 354 } 355 p = whereScanNext(&scan); 356 } 357 return pResult; 358 } 359 360 /* 361 ** This function searches pList for an entry that matches the iCol-th column 362 ** of index pIdx. 363 ** 364 ** If such an expression is found, its index in pList->a[] is returned. If 365 ** no expression is found, -1 is returned. 366 */ 367 static int findIndexCol( 368 Parse *pParse, /* Parse context */ 369 ExprList *pList, /* Expression list to search */ 370 int iBase, /* Cursor for table associated with pIdx */ 371 Index *pIdx, /* Index to match column of */ 372 int iCol /* Column of index to match */ 373 ){ 374 int i; 375 const char *zColl = pIdx->azColl[iCol]; 376 377 for(i=0; i<pList->nExpr; i++){ 378 Expr *p = sqlite3ExprSkipCollate(pList->a[i].pExpr); 379 if( p->op==TK_COLUMN 380 && p->iColumn==pIdx->aiColumn[iCol] 381 && p->iTable==iBase 382 ){ 383 CollSeq *pColl = sqlite3ExprCollSeq(pParse, pList->a[i].pExpr); 384 if( pColl && 0==sqlite3StrICmp(pColl->zName, zColl) ){ 385 return i; 386 } 387 } 388 } 389 390 return -1; 391 } 392 393 /* 394 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL 395 */ 396 static int indexColumnNotNull(Index *pIdx, int iCol){ 397 int j; 398 assert( pIdx!=0 ); 399 assert( iCol>=0 && iCol<pIdx->nColumn ); 400 j = pIdx->aiColumn[iCol]; 401 if( j>=0 ){ 402 return pIdx->pTable->aCol[j].notNull; 403 }else if( j==(-1) ){ 404 return 1; 405 }else{ 406 assert( j==(-2) ); 407 return 0; /* Assume an indexed expression can always yield a NULL */ 408 409 } 410 } 411 412 /* 413 ** Return true if the DISTINCT expression-list passed as the third argument 414 ** is redundant. 415 ** 416 ** A DISTINCT list is redundant if any subset of the columns in the 417 ** DISTINCT list are collectively unique and individually non-null. 418 */ 419 static int isDistinctRedundant( 420 Parse *pParse, /* Parsing context */ 421 SrcList *pTabList, /* The FROM clause */ 422 WhereClause *pWC, /* The WHERE clause */ 423 ExprList *pDistinct /* The result set that needs to be DISTINCT */ 424 ){ 425 Table *pTab; 426 Index *pIdx; 427 int i; 428 int iBase; 429 430 /* If there is more than one table or sub-select in the FROM clause of 431 ** this query, then it will not be possible to show that the DISTINCT 432 ** clause is redundant. */ 433 if( pTabList->nSrc!=1 ) return 0; 434 iBase = pTabList->a[0].iCursor; 435 pTab = pTabList->a[0].pTab; 436 437 /* If any of the expressions is an IPK column on table iBase, then return 438 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 439 ** current SELECT is a correlated sub-query. 440 */ 441 for(i=0; i<pDistinct->nExpr; i++){ 442 Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr); 443 if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1; 444 } 445 446 /* Loop through all indices on the table, checking each to see if it makes 447 ** the DISTINCT qualifier redundant. It does so if: 448 ** 449 ** 1. The index is itself UNIQUE, and 450 ** 451 ** 2. All of the columns in the index are either part of the pDistinct 452 ** list, or else the WHERE clause contains a term of the form "col=X", 453 ** where X is a constant value. The collation sequences of the 454 ** comparison and select-list expressions must match those of the index. 455 ** 456 ** 3. All of those index columns for which the WHERE clause does not 457 ** contain a "col=X" term are subject to a NOT NULL constraint. 458 */ 459 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 460 if( !IsUniqueIndex(pIdx) ) continue; 461 for(i=0; i<pIdx->nKeyCol; i++){ 462 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){ 463 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break; 464 if( indexColumnNotNull(pIdx, i)==0 ) break; 465 } 466 } 467 if( i==pIdx->nKeyCol ){ 468 /* This index implies that the DISTINCT qualifier is redundant. */ 469 return 1; 470 } 471 } 472 473 return 0; 474 } 475 476 477 /* 478 ** Estimate the logarithm of the input value to base 2. 479 */ 480 static LogEst estLog(LogEst N){ 481 return N<=10 ? 0 : sqlite3LogEst(N) - 33; 482 } 483 484 /* 485 ** Convert OP_Column opcodes to OP_Copy in previously generated code. 486 ** 487 ** This routine runs over generated VDBE code and translates OP_Column 488 ** opcodes into OP_Copy when the table is being accessed via co-routine 489 ** instead of via table lookup. 490 ** 491 ** If the bIncrRowid parameter is 0, then any OP_Rowid instructions on 492 ** cursor iTabCur are transformed into OP_Null. Or, if bIncrRowid is non-zero, 493 ** then each OP_Rowid is transformed into an instruction to increment the 494 ** value stored in its output register. 495 */ 496 static void translateColumnToCopy( 497 Vdbe *v, /* The VDBE containing code to translate */ 498 int iStart, /* Translate from this opcode to the end */ 499 int iTabCur, /* OP_Column/OP_Rowid references to this table */ 500 int iRegister, /* The first column is in this register */ 501 int bIncrRowid /* If non-zero, transform OP_rowid to OP_AddImm(1) */ 502 ){ 503 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart); 504 int iEnd = sqlite3VdbeCurrentAddr(v); 505 for(; iStart<iEnd; iStart++, pOp++){ 506 if( pOp->p1!=iTabCur ) continue; 507 if( pOp->opcode==OP_Column ){ 508 pOp->opcode = OP_Copy; 509 pOp->p1 = pOp->p2 + iRegister; 510 pOp->p2 = pOp->p3; 511 pOp->p3 = 0; 512 }else if( pOp->opcode==OP_Rowid ){ 513 if( bIncrRowid ){ 514 /* Increment the value stored in the P2 operand of the OP_Rowid. */ 515 pOp->opcode = OP_AddImm; 516 pOp->p1 = pOp->p2; 517 pOp->p2 = 1; 518 }else{ 519 pOp->opcode = OP_Null; 520 pOp->p1 = 0; 521 pOp->p3 = 0; 522 } 523 } 524 } 525 } 526 527 /* 528 ** Two routines for printing the content of an sqlite3_index_info 529 ** structure. Used for testing and debugging only. If neither 530 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 531 ** are no-ops. 532 */ 533 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) 534 static void TRACE_IDX_INPUTS(sqlite3_index_info *p){ 535 int i; 536 if( !sqlite3WhereTrace ) return; 537 for(i=0; i<p->nConstraint; i++){ 538 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", 539 i, 540 p->aConstraint[i].iColumn, 541 p->aConstraint[i].iTermOffset, 542 p->aConstraint[i].op, 543 p->aConstraint[i].usable); 544 } 545 for(i=0; i<p->nOrderBy; i++){ 546 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 547 i, 548 p->aOrderBy[i].iColumn, 549 p->aOrderBy[i].desc); 550 } 551 } 552 static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){ 553 int i; 554 if( !sqlite3WhereTrace ) return; 555 for(i=0; i<p->nConstraint; i++){ 556 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 557 i, 558 p->aConstraintUsage[i].argvIndex, 559 p->aConstraintUsage[i].omit); 560 } 561 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 562 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 563 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 564 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 565 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); 566 } 567 #else 568 #define TRACE_IDX_INPUTS(A) 569 #define TRACE_IDX_OUTPUTS(A) 570 #endif 571 572 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 573 /* 574 ** Return TRUE if the WHERE clause term pTerm is of a form where it 575 ** could be used with an index to access pSrc, assuming an appropriate 576 ** index existed. 577 */ 578 static int termCanDriveIndex( 579 WhereTerm *pTerm, /* WHERE clause term to check */ 580 struct SrcList_item *pSrc, /* Table we are trying to access */ 581 Bitmask notReady /* Tables in outer loops of the join */ 582 ){ 583 char aff; 584 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 585 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0; 586 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 587 if( pTerm->u.leftColumn<0 ) return 0; 588 aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity; 589 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 590 testcase( pTerm->pExpr->op==TK_IS ); 591 return 1; 592 } 593 #endif 594 595 596 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 597 /* 598 ** Generate code to construct the Index object for an automatic index 599 ** and to set up the WhereLevel object pLevel so that the code generator 600 ** makes use of the automatic index. 601 */ 602 static void constructAutomaticIndex( 603 Parse *pParse, /* The parsing context */ 604 WhereClause *pWC, /* The WHERE clause */ 605 struct SrcList_item *pSrc, /* The FROM clause term to get the next index */ 606 Bitmask notReady, /* Mask of cursors that are not available */ 607 WhereLevel *pLevel /* Write new index here */ 608 ){ 609 int nKeyCol; /* Number of columns in the constructed index */ 610 WhereTerm *pTerm; /* A single term of the WHERE clause */ 611 WhereTerm *pWCEnd; /* End of pWC->a[] */ 612 Index *pIdx; /* Object describing the transient index */ 613 Vdbe *v; /* Prepared statement under construction */ 614 int addrInit; /* Address of the initialization bypass jump */ 615 Table *pTable; /* The table being indexed */ 616 int addrTop; /* Top of the index fill loop */ 617 int regRecord; /* Register holding an index record */ 618 int n; /* Column counter */ 619 int i; /* Loop counter */ 620 int mxBitCol; /* Maximum column in pSrc->colUsed */ 621 CollSeq *pColl; /* Collating sequence to on a column */ 622 WhereLoop *pLoop; /* The Loop object */ 623 char *zNotUsed; /* Extra space on the end of pIdx */ 624 Bitmask idxCols; /* Bitmap of columns used for indexing */ 625 Bitmask extraCols; /* Bitmap of additional columns */ 626 u8 sentWarning = 0; /* True if a warnning has been issued */ 627 Expr *pPartial = 0; /* Partial Index Expression */ 628 int iContinue = 0; /* Jump here to skip excluded rows */ 629 struct SrcList_item *pTabItem; /* FROM clause term being indexed */ 630 int addrCounter = 0; /* Address where integer counter is initialized */ 631 int regBase; /* Array of registers where record is assembled */ 632 633 /* Generate code to skip over the creation and initialization of the 634 ** transient index on 2nd and subsequent iterations of the loop. */ 635 v = pParse->pVdbe; 636 assert( v!=0 ); 637 addrInit = sqlite3CodeOnce(pParse); VdbeCoverage(v); 638 639 /* Count the number of columns that will be added to the index 640 ** and used to match WHERE clause constraints */ 641 nKeyCol = 0; 642 pTable = pSrc->pTab; 643 pWCEnd = &pWC->a[pWC->nTerm]; 644 pLoop = pLevel->pWLoop; 645 idxCols = 0; 646 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 647 Expr *pExpr = pTerm->pExpr; 648 assert( !ExprHasProperty(pExpr, EP_FromJoin) /* prereq always non-zero */ 649 || pExpr->iRightJoinTable!=pSrc->iCursor /* for the right-hand */ 650 || pLoop->prereq!=0 ); /* table of a LEFT JOIN */ 651 if( pLoop->prereq==0 652 && (pTerm->wtFlags & TERM_VIRTUAL)==0 653 && !ExprHasProperty(pExpr, EP_FromJoin) 654 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) ){ 655 pPartial = sqlite3ExprAnd(pParse->db, pPartial, 656 sqlite3ExprDup(pParse->db, pExpr, 0)); 657 } 658 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 659 int iCol = pTerm->u.leftColumn; 660 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 661 testcase( iCol==BMS ); 662 testcase( iCol==BMS-1 ); 663 if( !sentWarning ){ 664 sqlite3_log(SQLITE_WARNING_AUTOINDEX, 665 "automatic index on %s(%s)", pTable->zName, 666 pTable->aCol[iCol].zName); 667 sentWarning = 1; 668 } 669 if( (idxCols & cMask)==0 ){ 670 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ 671 goto end_auto_index_create; 672 } 673 pLoop->aLTerm[nKeyCol++] = pTerm; 674 idxCols |= cMask; 675 } 676 } 677 } 678 assert( nKeyCol>0 ); 679 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; 680 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED 681 | WHERE_AUTO_INDEX; 682 683 /* Count the number of additional columns needed to create a 684 ** covering index. A "covering index" is an index that contains all 685 ** columns that are needed by the query. With a covering index, the 686 ** original table never needs to be accessed. Automatic indices must 687 ** be a covering index because the index will not be updated if the 688 ** original table changes and the index and table cannot both be used 689 ** if they go out of sync. 690 */ 691 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); 692 mxBitCol = MIN(BMS-1,pTable->nCol); 693 testcase( pTable->nCol==BMS-1 ); 694 testcase( pTable->nCol==BMS-2 ); 695 for(i=0; i<mxBitCol; i++){ 696 if( extraCols & MASKBIT(i) ) nKeyCol++; 697 } 698 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 699 nKeyCol += pTable->nCol - BMS + 1; 700 } 701 702 /* Construct the Index object to describe this index */ 703 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); 704 if( pIdx==0 ) goto end_auto_index_create; 705 pLoop->u.btree.pIndex = pIdx; 706 pIdx->zName = "auto-index"; 707 pIdx->pTable = pTable; 708 n = 0; 709 idxCols = 0; 710 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 711 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 712 int iCol = pTerm->u.leftColumn; 713 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 714 testcase( iCol==BMS-1 ); 715 testcase( iCol==BMS ); 716 if( (idxCols & cMask)==0 ){ 717 Expr *pX = pTerm->pExpr; 718 idxCols |= cMask; 719 pIdx->aiColumn[n] = pTerm->u.leftColumn; 720 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); 721 pIdx->azColl[n] = pColl ? pColl->zName : "BINARY"; 722 n++; 723 } 724 } 725 } 726 assert( (u32)n==pLoop->u.btree.nEq ); 727 728 /* Add additional columns needed to make the automatic index into 729 ** a covering index */ 730 for(i=0; i<mxBitCol; i++){ 731 if( extraCols & MASKBIT(i) ){ 732 pIdx->aiColumn[n] = i; 733 pIdx->azColl[n] = "BINARY"; 734 n++; 735 } 736 } 737 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 738 for(i=BMS-1; i<pTable->nCol; i++){ 739 pIdx->aiColumn[n] = i; 740 pIdx->azColl[n] = "BINARY"; 741 n++; 742 } 743 } 744 assert( n==nKeyCol ); 745 pIdx->aiColumn[n] = XN_ROWID; 746 pIdx->azColl[n] = "BINARY"; 747 748 /* Create the automatic index */ 749 assert( pLevel->iIdxCur>=0 ); 750 pLevel->iIdxCur = pParse->nTab++; 751 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); 752 sqlite3VdbeSetP4KeyInfo(pParse, pIdx); 753 VdbeComment((v, "for %s", pTable->zName)); 754 755 /* Fill the automatic index with content */ 756 sqlite3ExprCachePush(pParse); 757 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom]; 758 if( pTabItem->fg.viaCoroutine ){ 759 int regYield = pTabItem->regReturn; 760 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0); 761 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); 762 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 763 VdbeCoverage(v); 764 VdbeComment((v, "next row of \"%s\"", pTabItem->pTab->zName)); 765 }else{ 766 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); 767 } 768 if( pPartial ){ 769 iContinue = sqlite3VdbeMakeLabel(v); 770 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); 771 pLoop->wsFlags |= WHERE_PARTIALIDX; 772 } 773 regRecord = sqlite3GetTempReg(pParse); 774 regBase = sqlite3GenerateIndexKey( 775 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0 776 ); 777 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 778 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 779 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); 780 if( pTabItem->fg.viaCoroutine ){ 781 sqlite3VdbeChangeP2(v, addrCounter, regBase+n); 782 translateColumnToCopy(v, addrTop, pLevel->iTabCur, pTabItem->regResult, 1); 783 sqlite3VdbeGoto(v, addrTop); 784 pTabItem->fg.viaCoroutine = 0; 785 }else{ 786 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); 787 } 788 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 789 sqlite3VdbeJumpHere(v, addrTop); 790 sqlite3ReleaseTempReg(pParse, regRecord); 791 sqlite3ExprCachePop(pParse); 792 793 /* Jump here when skipping the initialization */ 794 sqlite3VdbeJumpHere(v, addrInit); 795 796 end_auto_index_create: 797 sqlite3ExprDelete(pParse->db, pPartial); 798 } 799 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 800 801 #ifndef SQLITE_OMIT_VIRTUALTABLE 802 /* 803 ** Allocate and populate an sqlite3_index_info structure. It is the 804 ** responsibility of the caller to eventually release the structure 805 ** by passing the pointer returned by this function to sqlite3_free(). 806 */ 807 static sqlite3_index_info *allocateIndexInfo( 808 Parse *pParse, 809 WhereClause *pWC, 810 Bitmask mUnusable, /* Ignore terms with these prereqs */ 811 struct SrcList_item *pSrc, 812 ExprList *pOrderBy 813 ){ 814 int i, j; 815 int nTerm; 816 struct sqlite3_index_constraint *pIdxCons; 817 struct sqlite3_index_orderby *pIdxOrderBy; 818 struct sqlite3_index_constraint_usage *pUsage; 819 WhereTerm *pTerm; 820 int nOrderBy; 821 sqlite3_index_info *pIdxInfo; 822 823 /* Count the number of possible WHERE clause constraints referring 824 ** to this virtual table */ 825 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 826 if( pTerm->leftCursor != pSrc->iCursor ) continue; 827 if( pTerm->prereqRight & mUnusable ) continue; 828 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 829 testcase( pTerm->eOperator & WO_IN ); 830 testcase( pTerm->eOperator & WO_ISNULL ); 831 testcase( pTerm->eOperator & WO_IS ); 832 testcase( pTerm->eOperator & WO_ALL ); 833 if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV|WO_IS))==0 ) continue; 834 if( pTerm->wtFlags & TERM_VNULL ) continue; 835 assert( pTerm->u.leftColumn>=(-1) ); 836 nTerm++; 837 } 838 839 /* If the ORDER BY clause contains only columns in the current 840 ** virtual table then allocate space for the aOrderBy part of 841 ** the sqlite3_index_info structure. 842 */ 843 nOrderBy = 0; 844 if( pOrderBy ){ 845 int n = pOrderBy->nExpr; 846 for(i=0; i<n; i++){ 847 Expr *pExpr = pOrderBy->a[i].pExpr; 848 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; 849 } 850 if( i==n){ 851 nOrderBy = n; 852 } 853 } 854 855 /* Allocate the sqlite3_index_info structure 856 */ 857 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 858 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 859 + sizeof(*pIdxOrderBy)*nOrderBy ); 860 if( pIdxInfo==0 ){ 861 sqlite3ErrorMsg(pParse, "out of memory"); 862 return 0; 863 } 864 865 /* Initialize the structure. The sqlite3_index_info structure contains 866 ** many fields that are declared "const" to prevent xBestIndex from 867 ** changing them. We have to do some funky casting in order to 868 ** initialize those fields. 869 */ 870 pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1]; 871 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 872 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 873 *(int*)&pIdxInfo->nConstraint = nTerm; 874 *(int*)&pIdxInfo->nOrderBy = nOrderBy; 875 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons; 876 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy; 877 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage = 878 pUsage; 879 880 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 881 u8 op; 882 if( pTerm->leftCursor != pSrc->iCursor ) continue; 883 if( pTerm->prereqRight & mUnusable ) continue; 884 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 885 testcase( pTerm->eOperator & WO_IN ); 886 testcase( pTerm->eOperator & WO_IS ); 887 testcase( pTerm->eOperator & WO_ISNULL ); 888 testcase( pTerm->eOperator & WO_ALL ); 889 if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV|WO_IS))==0 ) continue; 890 if( pTerm->wtFlags & TERM_VNULL ) continue; 891 assert( pTerm->u.leftColumn>=(-1) ); 892 pIdxCons[j].iColumn = pTerm->u.leftColumn; 893 pIdxCons[j].iTermOffset = i; 894 op = (u8)pTerm->eOperator & WO_ALL; 895 if( op==WO_IN ) op = WO_EQ; 896 pIdxCons[j].op = op; 897 /* The direct assignment in the previous line is possible only because 898 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 899 ** following asserts verify this fact. */ 900 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 901 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 902 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 903 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 904 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 905 assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH ); 906 assert( pTerm->eOperator & (WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) ); 907 j++; 908 } 909 for(i=0; i<nOrderBy; i++){ 910 Expr *pExpr = pOrderBy->a[i].pExpr; 911 pIdxOrderBy[i].iColumn = pExpr->iColumn; 912 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder; 913 } 914 915 return pIdxInfo; 916 } 917 918 /* 919 ** The table object reference passed as the second argument to this function 920 ** must represent a virtual table. This function invokes the xBestIndex() 921 ** method of the virtual table with the sqlite3_index_info object that 922 ** comes in as the 3rd argument to this function. 923 ** 924 ** If an error occurs, pParse is populated with an error message and a 925 ** non-zero value is returned. Otherwise, 0 is returned and the output 926 ** part of the sqlite3_index_info structure is left populated. 927 ** 928 ** Whether or not an error is returned, it is the responsibility of the 929 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 930 ** that this is required. 931 */ 932 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 933 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 934 int i; 935 int rc; 936 937 TRACE_IDX_INPUTS(p); 938 rc = pVtab->pModule->xBestIndex(pVtab, p); 939 TRACE_IDX_OUTPUTS(p); 940 941 if( rc!=SQLITE_OK ){ 942 if( rc==SQLITE_NOMEM ){ 943 pParse->db->mallocFailed = 1; 944 }else if( !pVtab->zErrMsg ){ 945 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 946 }else{ 947 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 948 } 949 } 950 sqlite3_free(pVtab->zErrMsg); 951 pVtab->zErrMsg = 0; 952 953 for(i=0; i<p->nConstraint; i++){ 954 if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){ 955 sqlite3ErrorMsg(pParse, 956 "table %s: xBestIndex returned an invalid plan", pTab->zName); 957 } 958 } 959 960 return pParse->nErr; 961 } 962 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 963 964 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 965 /* 966 ** Estimate the location of a particular key among all keys in an 967 ** index. Store the results in aStat as follows: 968 ** 969 ** aStat[0] Est. number of rows less than pRec 970 ** aStat[1] Est. number of rows equal to pRec 971 ** 972 ** Return the index of the sample that is the smallest sample that 973 ** is greater than or equal to pRec. Note that this index is not an index 974 ** into the aSample[] array - it is an index into a virtual set of samples 975 ** based on the contents of aSample[] and the number of fields in record 976 ** pRec. 977 */ 978 static int whereKeyStats( 979 Parse *pParse, /* Database connection */ 980 Index *pIdx, /* Index to consider domain of */ 981 UnpackedRecord *pRec, /* Vector of values to consider */ 982 int roundUp, /* Round up if true. Round down if false */ 983 tRowcnt *aStat /* OUT: stats written here */ 984 ){ 985 IndexSample *aSample = pIdx->aSample; 986 int iCol; /* Index of required stats in anEq[] etc. */ 987 int i; /* Index of first sample >= pRec */ 988 int iSample; /* Smallest sample larger than or equal to pRec */ 989 int iMin = 0; /* Smallest sample not yet tested */ 990 int iTest; /* Next sample to test */ 991 int res; /* Result of comparison operation */ 992 int nField; /* Number of fields in pRec */ 993 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 994 995 #ifndef SQLITE_DEBUG 996 UNUSED_PARAMETER( pParse ); 997 #endif 998 assert( pRec!=0 ); 999 assert( pIdx->nSample>0 ); 1000 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1001 1002 /* Do a binary search to find the first sample greater than or equal 1003 ** to pRec. If pRec contains a single field, the set of samples to search 1004 ** is simply the aSample[] array. If the samples in aSample[] contain more 1005 ** than one fields, all fields following the first are ignored. 1006 ** 1007 ** If pRec contains N fields, where N is more than one, then as well as the 1008 ** samples in aSample[] (truncated to N fields), the search also has to 1009 ** consider prefixes of those samples. For example, if the set of samples 1010 ** in aSample is: 1011 ** 1012 ** aSample[0] = (a, 5) 1013 ** aSample[1] = (a, 10) 1014 ** aSample[2] = (b, 5) 1015 ** aSample[3] = (c, 100) 1016 ** aSample[4] = (c, 105) 1017 ** 1018 ** Then the search space should ideally be the samples above and the 1019 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1020 ** the code actually searches this set: 1021 ** 1022 ** 0: (a) 1023 ** 1: (a, 5) 1024 ** 2: (a, 10) 1025 ** 3: (a, 10) 1026 ** 4: (b) 1027 ** 5: (b, 5) 1028 ** 6: (c) 1029 ** 7: (c, 100) 1030 ** 8: (c, 105) 1031 ** 9: (c, 105) 1032 ** 1033 ** For each sample in the aSample[] array, N samples are present in the 1034 ** effective sample array. In the above, samples 0 and 1 are based on 1035 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1036 ** 1037 ** Often, sample i of each block of N effective samples has (i+1) fields. 1038 ** Except, each sample may be extended to ensure that it is greater than or 1039 ** equal to the previous sample in the array. For example, in the above, 1040 ** sample 2 is the first sample of a block of N samples, so at first it 1041 ** appears that it should be 1 field in size. However, that would make it 1042 ** smaller than sample 1, so the binary search would not work. As a result, 1043 ** it is extended to two fields. The duplicates that this creates do not 1044 ** cause any problems. 1045 */ 1046 nField = pRec->nField; 1047 iCol = 0; 1048 iSample = pIdx->nSample * nField; 1049 do{ 1050 int iSamp; /* Index in aSample[] of test sample */ 1051 int n; /* Number of fields in test sample */ 1052 1053 iTest = (iMin+iSample)/2; 1054 iSamp = iTest / nField; 1055 if( iSamp>0 ){ 1056 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1057 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1058 ** fields that is greater than the previous effective sample. */ 1059 for(n=(iTest % nField) + 1; n<nField; n++){ 1060 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1061 } 1062 }else{ 1063 n = iTest + 1; 1064 } 1065 1066 pRec->nField = n; 1067 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1068 if( res<0 ){ 1069 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1070 iMin = iTest+1; 1071 }else if( res==0 && n<nField ){ 1072 iLower = aSample[iSamp].anLt[n-1]; 1073 iMin = iTest+1; 1074 res = -1; 1075 }else{ 1076 iSample = iTest; 1077 iCol = n-1; 1078 } 1079 }while( res && iMin<iSample ); 1080 i = iSample / nField; 1081 1082 #ifdef SQLITE_DEBUG 1083 /* The following assert statements check that the binary search code 1084 ** above found the right answer. This block serves no purpose other 1085 ** than to invoke the asserts. */ 1086 if( pParse->db->mallocFailed==0 ){ 1087 if( res==0 ){ 1088 /* If (res==0) is true, then pRec must be equal to sample i. */ 1089 assert( i<pIdx->nSample ); 1090 assert( iCol==nField-1 ); 1091 pRec->nField = nField; 1092 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1093 || pParse->db->mallocFailed 1094 ); 1095 }else{ 1096 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1097 ** all samples in the aSample[] array, pRec must be smaller than the 1098 ** (iCol+1) field prefix of sample i. */ 1099 assert( i<=pIdx->nSample && i>=0 ); 1100 pRec->nField = iCol+1; 1101 assert( i==pIdx->nSample 1102 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1103 || pParse->db->mallocFailed ); 1104 1105 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1106 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1107 ** be greater than or equal to the (iCol) field prefix of sample i. 1108 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1109 if( iCol>0 ){ 1110 pRec->nField = iCol; 1111 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1112 || pParse->db->mallocFailed ); 1113 } 1114 if( i>0 ){ 1115 pRec->nField = nField; 1116 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1117 || pParse->db->mallocFailed ); 1118 } 1119 } 1120 } 1121 #endif /* ifdef SQLITE_DEBUG */ 1122 1123 if( res==0 ){ 1124 /* Record pRec is equal to sample i */ 1125 assert( iCol==nField-1 ); 1126 aStat[0] = aSample[i].anLt[iCol]; 1127 aStat[1] = aSample[i].anEq[iCol]; 1128 }else{ 1129 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1130 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1131 ** is larger than all samples in the array. */ 1132 tRowcnt iUpper, iGap; 1133 if( i>=pIdx->nSample ){ 1134 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1135 }else{ 1136 iUpper = aSample[i].anLt[iCol]; 1137 } 1138 1139 if( iLower>=iUpper ){ 1140 iGap = 0; 1141 }else{ 1142 iGap = iUpper - iLower; 1143 } 1144 if( roundUp ){ 1145 iGap = (iGap*2)/3; 1146 }else{ 1147 iGap = iGap/3; 1148 } 1149 aStat[0] = iLower + iGap; 1150 aStat[1] = pIdx->aAvgEq[iCol]; 1151 } 1152 1153 /* Restore the pRec->nField value before returning. */ 1154 pRec->nField = nField; 1155 return i; 1156 } 1157 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1158 1159 /* 1160 ** If it is not NULL, pTerm is a term that provides an upper or lower 1161 ** bound on a range scan. Without considering pTerm, it is estimated 1162 ** that the scan will visit nNew rows. This function returns the number 1163 ** estimated to be visited after taking pTerm into account. 1164 ** 1165 ** If the user explicitly specified a likelihood() value for this term, 1166 ** then the return value is the likelihood multiplied by the number of 1167 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1168 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1169 */ 1170 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1171 LogEst nRet = nNew; 1172 if( pTerm ){ 1173 if( pTerm->truthProb<=0 ){ 1174 nRet += pTerm->truthProb; 1175 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1176 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1177 } 1178 } 1179 return nRet; 1180 } 1181 1182 1183 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1184 /* 1185 ** Return the affinity for a single column of an index. 1186 */ 1187 static char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1188 assert( iCol>=0 && iCol<pIdx->nColumn ); 1189 if( !pIdx->zColAff ){ 1190 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1191 } 1192 return pIdx->zColAff[iCol]; 1193 } 1194 #endif 1195 1196 1197 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1198 /* 1199 ** This function is called to estimate the number of rows visited by a 1200 ** range-scan on a skip-scan index. For example: 1201 ** 1202 ** CREATE INDEX i1 ON t1(a, b, c); 1203 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1204 ** 1205 ** Value pLoop->nOut is currently set to the estimated number of rows 1206 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1207 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1208 ** on the stat4 data for the index. this scan will be peformed multiple 1209 ** times (once for each (a,b) combination that matches a=?) is dealt with 1210 ** by the caller. 1211 ** 1212 ** It does this by scanning through all stat4 samples, comparing values 1213 ** extracted from pLower and pUpper with the corresponding column in each 1214 ** sample. If L and U are the number of samples found to be less than or 1215 ** equal to the values extracted from pLower and pUpper respectively, and 1216 ** N is the total number of samples, the pLoop->nOut value is adjusted 1217 ** as follows: 1218 ** 1219 ** nOut = nOut * ( min(U - L, 1) / N ) 1220 ** 1221 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1222 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1223 ** U is set to N. 1224 ** 1225 ** Normally, this function sets *pbDone to 1 before returning. However, 1226 ** if no value can be extracted from either pLower or pUpper (and so the 1227 ** estimate of the number of rows delivered remains unchanged), *pbDone 1228 ** is left as is. 1229 ** 1230 ** If an error occurs, an SQLite error code is returned. Otherwise, 1231 ** SQLITE_OK. 1232 */ 1233 static int whereRangeSkipScanEst( 1234 Parse *pParse, /* Parsing & code generating context */ 1235 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1236 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1237 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1238 int *pbDone /* Set to true if at least one expr. value extracted */ 1239 ){ 1240 Index *p = pLoop->u.btree.pIndex; 1241 int nEq = pLoop->u.btree.nEq; 1242 sqlite3 *db = pParse->db; 1243 int nLower = -1; 1244 int nUpper = p->nSample+1; 1245 int rc = SQLITE_OK; 1246 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1247 CollSeq *pColl; 1248 1249 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1250 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1251 sqlite3_value *pVal = 0; /* Value extracted from record */ 1252 1253 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1254 if( pLower ){ 1255 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1256 nLower = 0; 1257 } 1258 if( pUpper && rc==SQLITE_OK ){ 1259 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1260 nUpper = p2 ? 0 : p->nSample; 1261 } 1262 1263 if( p1 || p2 ){ 1264 int i; 1265 int nDiff; 1266 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1267 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1268 if( rc==SQLITE_OK && p1 ){ 1269 int res = sqlite3MemCompare(p1, pVal, pColl); 1270 if( res>=0 ) nLower++; 1271 } 1272 if( rc==SQLITE_OK && p2 ){ 1273 int res = sqlite3MemCompare(p2, pVal, pColl); 1274 if( res>=0 ) nUpper++; 1275 } 1276 } 1277 nDiff = (nUpper - nLower); 1278 if( nDiff<=0 ) nDiff = 1; 1279 1280 /* If there is both an upper and lower bound specified, and the 1281 ** comparisons indicate that they are close together, use the fallback 1282 ** method (assume that the scan visits 1/64 of the rows) for estimating 1283 ** the number of rows visited. Otherwise, estimate the number of rows 1284 ** using the method described in the header comment for this function. */ 1285 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1286 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1287 pLoop->nOut -= nAdjust; 1288 *pbDone = 1; 1289 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1290 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1291 } 1292 1293 }else{ 1294 assert( *pbDone==0 ); 1295 } 1296 1297 sqlite3ValueFree(p1); 1298 sqlite3ValueFree(p2); 1299 sqlite3ValueFree(pVal); 1300 1301 return rc; 1302 } 1303 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1304 1305 /* 1306 ** This function is used to estimate the number of rows that will be visited 1307 ** by scanning an index for a range of values. The range may have an upper 1308 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1309 ** and lower bounds are represented by pLower and pUpper respectively. For 1310 ** example, assuming that index p is on t1(a): 1311 ** 1312 ** ... FROM t1 WHERE a > ? AND a < ? ... 1313 ** |_____| |_____| 1314 ** | | 1315 ** pLower pUpper 1316 ** 1317 ** If either of the upper or lower bound is not present, then NULL is passed in 1318 ** place of the corresponding WhereTerm. 1319 ** 1320 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1321 ** column subject to the range constraint. Or, equivalently, the number of 1322 ** equality constraints optimized by the proposed index scan. For example, 1323 ** assuming index p is on t1(a, b), and the SQL query is: 1324 ** 1325 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1326 ** 1327 ** then nEq is set to 1 (as the range restricted column, b, is the second 1328 ** left-most column of the index). Or, if the query is: 1329 ** 1330 ** ... FROM t1 WHERE a > ? AND a < ? ... 1331 ** 1332 ** then nEq is set to 0. 1333 ** 1334 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1335 ** number of rows that the index scan is expected to visit without 1336 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1337 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1338 ** to account for the range constraints pLower and pUpper. 1339 ** 1340 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1341 ** used, a single range inequality reduces the search space by a factor of 4. 1342 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1343 ** rows visited by a factor of 64. 1344 */ 1345 static int whereRangeScanEst( 1346 Parse *pParse, /* Parsing & code generating context */ 1347 WhereLoopBuilder *pBuilder, 1348 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1349 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1350 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1351 ){ 1352 int rc = SQLITE_OK; 1353 int nOut = pLoop->nOut; 1354 LogEst nNew; 1355 1356 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1357 Index *p = pLoop->u.btree.pIndex; 1358 int nEq = pLoop->u.btree.nEq; 1359 1360 if( p->nSample>0 && nEq<p->nSampleCol ){ 1361 if( nEq==pBuilder->nRecValid ){ 1362 UnpackedRecord *pRec = pBuilder->pRec; 1363 tRowcnt a[2]; 1364 u8 aff; 1365 1366 /* Variable iLower will be set to the estimate of the number of rows in 1367 ** the index that are less than the lower bound of the range query. The 1368 ** lower bound being the concatenation of $P and $L, where $P is the 1369 ** key-prefix formed by the nEq values matched against the nEq left-most 1370 ** columns of the index, and $L is the value in pLower. 1371 ** 1372 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1373 ** is not a simple variable or literal value), the lower bound of the 1374 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1375 ** if $L is available, whereKeyStats() is called for both ($P) and 1376 ** ($P:$L) and the larger of the two returned values is used. 1377 ** 1378 ** Similarly, iUpper is to be set to the estimate of the number of rows 1379 ** less than the upper bound of the range query. Where the upper bound 1380 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1381 ** of iUpper are requested of whereKeyStats() and the smaller used. 1382 ** 1383 ** The number of rows between the two bounds is then just iUpper-iLower. 1384 */ 1385 tRowcnt iLower; /* Rows less than the lower bound */ 1386 tRowcnt iUpper; /* Rows less than the upper bound */ 1387 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1388 int iUprIdx = -1; /* aSample[] for the upper bound */ 1389 1390 if( pRec ){ 1391 testcase( pRec->nField!=pBuilder->nRecValid ); 1392 pRec->nField = pBuilder->nRecValid; 1393 } 1394 aff = sqlite3IndexColumnAffinity(pParse->db, p, nEq); 1395 assert( nEq!=p->nKeyCol || aff==SQLITE_AFF_INTEGER ); 1396 /* Determine iLower and iUpper using ($P) only. */ 1397 if( nEq==0 ){ 1398 iLower = 0; 1399 iUpper = p->nRowEst0; 1400 }else{ 1401 /* Note: this call could be optimized away - since the same values must 1402 ** have been requested when testing key $P in whereEqualScanEst(). */ 1403 whereKeyStats(pParse, p, pRec, 0, a); 1404 iLower = a[0]; 1405 iUpper = a[0] + a[1]; 1406 } 1407 1408 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1409 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1410 assert( p->aSortOrder!=0 ); 1411 if( p->aSortOrder[nEq] ){ 1412 /* The roles of pLower and pUpper are swapped for a DESC index */ 1413 SWAP(WhereTerm*, pLower, pUpper); 1414 } 1415 1416 /* If possible, improve on the iLower estimate using ($P:$L). */ 1417 if( pLower ){ 1418 int bOk; /* True if value is extracted from pExpr */ 1419 Expr *pExpr = pLower->pExpr->pRight; 1420 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq, &bOk); 1421 if( rc==SQLITE_OK && bOk ){ 1422 tRowcnt iNew; 1423 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1424 iNew = a[0] + ((pLower->eOperator & (WO_GT|WO_LE)) ? a[1] : 0); 1425 if( iNew>iLower ) iLower = iNew; 1426 nOut--; 1427 pLower = 0; 1428 } 1429 } 1430 1431 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1432 if( pUpper ){ 1433 int bOk; /* True if value is extracted from pExpr */ 1434 Expr *pExpr = pUpper->pExpr->pRight; 1435 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq, &bOk); 1436 if( rc==SQLITE_OK && bOk ){ 1437 tRowcnt iNew; 1438 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1439 iNew = a[0] + ((pUpper->eOperator & (WO_GT|WO_LE)) ? a[1] : 0); 1440 if( iNew<iUpper ) iUpper = iNew; 1441 nOut--; 1442 pUpper = 0; 1443 } 1444 } 1445 1446 pBuilder->pRec = pRec; 1447 if( rc==SQLITE_OK ){ 1448 if( iUpper>iLower ){ 1449 nNew = sqlite3LogEst(iUpper - iLower); 1450 /* TUNING: If both iUpper and iLower are derived from the same 1451 ** sample, then assume they are 4x more selective. This brings 1452 ** the estimated selectivity more in line with what it would be 1453 ** if estimated without the use of STAT3/4 tables. */ 1454 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1455 }else{ 1456 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1457 } 1458 if( nNew<nOut ){ 1459 nOut = nNew; 1460 } 1461 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1462 (u32)iLower, (u32)iUpper, nOut)); 1463 } 1464 }else{ 1465 int bDone = 0; 1466 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1467 if( bDone ) return rc; 1468 } 1469 } 1470 #else 1471 UNUSED_PARAMETER(pParse); 1472 UNUSED_PARAMETER(pBuilder); 1473 assert( pLower || pUpper ); 1474 #endif 1475 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1476 nNew = whereRangeAdjust(pLower, nOut); 1477 nNew = whereRangeAdjust(pUpper, nNew); 1478 1479 /* TUNING: If there is both an upper and lower limit and neither limit 1480 ** has an application-defined likelihood(), assume the range is 1481 ** reduced by an additional 75%. This means that, by default, an open-ended 1482 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1483 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1484 ** match 1/64 of the index. */ 1485 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1486 nNew -= 20; 1487 } 1488 1489 nOut -= (pLower!=0) + (pUpper!=0); 1490 if( nNew<10 ) nNew = 10; 1491 if( nNew<nOut ) nOut = nNew; 1492 #if defined(WHERETRACE_ENABLED) 1493 if( pLoop->nOut>nOut ){ 1494 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1495 pLoop->nOut, nOut)); 1496 } 1497 #endif 1498 pLoop->nOut = (LogEst)nOut; 1499 return rc; 1500 } 1501 1502 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1503 /* 1504 ** Estimate the number of rows that will be returned based on 1505 ** an equality constraint x=VALUE and where that VALUE occurs in 1506 ** the histogram data. This only works when x is the left-most 1507 ** column of an index and sqlite_stat3 histogram data is available 1508 ** for that index. When pExpr==NULL that means the constraint is 1509 ** "x IS NULL" instead of "x=VALUE". 1510 ** 1511 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1512 ** If unable to make an estimate, leave *pnRow unchanged and return 1513 ** non-zero. 1514 ** 1515 ** This routine can fail if it is unable to load a collating sequence 1516 ** required for string comparison, or if unable to allocate memory 1517 ** for a UTF conversion required for comparison. The error is stored 1518 ** in the pParse structure. 1519 */ 1520 static int whereEqualScanEst( 1521 Parse *pParse, /* Parsing & code generating context */ 1522 WhereLoopBuilder *pBuilder, 1523 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1524 tRowcnt *pnRow /* Write the revised row estimate here */ 1525 ){ 1526 Index *p = pBuilder->pNew->u.btree.pIndex; 1527 int nEq = pBuilder->pNew->u.btree.nEq; 1528 UnpackedRecord *pRec = pBuilder->pRec; 1529 u8 aff; /* Column affinity */ 1530 int rc; /* Subfunction return code */ 1531 tRowcnt a[2]; /* Statistics */ 1532 int bOk; 1533 1534 assert( nEq>=1 ); 1535 assert( nEq<=p->nColumn ); 1536 assert( p->aSample!=0 ); 1537 assert( p->nSample>0 ); 1538 assert( pBuilder->nRecValid<nEq ); 1539 1540 /* If values are not available for all fields of the index to the left 1541 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1542 if( pBuilder->nRecValid<(nEq-1) ){ 1543 return SQLITE_NOTFOUND; 1544 } 1545 1546 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1547 ** below would return the same value. */ 1548 if( nEq>=p->nColumn ){ 1549 *pnRow = 1; 1550 return SQLITE_OK; 1551 } 1552 1553 aff = sqlite3IndexColumnAffinity(pParse->db, p, nEq-1); 1554 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq-1, &bOk); 1555 pBuilder->pRec = pRec; 1556 if( rc!=SQLITE_OK ) return rc; 1557 if( bOk==0 ) return SQLITE_NOTFOUND; 1558 pBuilder->nRecValid = nEq; 1559 1560 whereKeyStats(pParse, p, pRec, 0, a); 1561 WHERETRACE(0x10,("equality scan regions: %d\n", (int)a[1])); 1562 *pnRow = a[1]; 1563 1564 return rc; 1565 } 1566 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1567 1568 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1569 /* 1570 ** Estimate the number of rows that will be returned based on 1571 ** an IN constraint where the right-hand side of the IN operator 1572 ** is a list of values. Example: 1573 ** 1574 ** WHERE x IN (1,2,3,4) 1575 ** 1576 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1577 ** If unable to make an estimate, leave *pnRow unchanged and return 1578 ** non-zero. 1579 ** 1580 ** This routine can fail if it is unable to load a collating sequence 1581 ** required for string comparison, or if unable to allocate memory 1582 ** for a UTF conversion required for comparison. The error is stored 1583 ** in the pParse structure. 1584 */ 1585 static int whereInScanEst( 1586 Parse *pParse, /* Parsing & code generating context */ 1587 WhereLoopBuilder *pBuilder, 1588 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1589 tRowcnt *pnRow /* Write the revised row estimate here */ 1590 ){ 1591 Index *p = pBuilder->pNew->u.btree.pIndex; 1592 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1593 int nRecValid = pBuilder->nRecValid; 1594 int rc = SQLITE_OK; /* Subfunction return code */ 1595 tRowcnt nEst; /* Number of rows for a single term */ 1596 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1597 int i; /* Loop counter */ 1598 1599 assert( p->aSample!=0 ); 1600 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1601 nEst = nRow0; 1602 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1603 nRowEst += nEst; 1604 pBuilder->nRecValid = nRecValid; 1605 } 1606 1607 if( rc==SQLITE_OK ){ 1608 if( nRowEst > nRow0 ) nRowEst = nRow0; 1609 *pnRow = nRowEst; 1610 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1611 } 1612 assert( pBuilder->nRecValid==nRecValid ); 1613 return rc; 1614 } 1615 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1616 1617 1618 #ifdef WHERETRACE_ENABLED 1619 /* 1620 ** Print the content of a WhereTerm object 1621 */ 1622 static void whereTermPrint(WhereTerm *pTerm, int iTerm){ 1623 if( pTerm==0 ){ 1624 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1625 }else{ 1626 char zType[4]; 1627 memcpy(zType, "...", 4); 1628 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1629 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1630 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1631 sqlite3DebugPrintf( 1632 "TERM-%-3d %p %s cursor=%-3d prob=%-3d op=0x%03x wtFlags=0x%04x\n", 1633 iTerm, pTerm, zType, pTerm->leftCursor, pTerm->truthProb, 1634 pTerm->eOperator, pTerm->wtFlags); 1635 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 1636 } 1637 } 1638 #endif 1639 1640 #ifdef WHERETRACE_ENABLED 1641 /* 1642 ** Print a WhereLoop object for debugging purposes 1643 */ 1644 static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){ 1645 WhereInfo *pWInfo = pWC->pWInfo; 1646 int nb = 1+(pWInfo->pTabList->nSrc+7)/8; 1647 struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab; 1648 Table *pTab = pItem->pTab; 1649 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 1650 p->iTab, nb, p->maskSelf, nb, p->prereq); 1651 sqlite3DebugPrintf(" %12s", 1652 pItem->zAlias ? pItem->zAlias : pTab->zName); 1653 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 1654 const char *zName; 1655 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 1656 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 1657 int i = sqlite3Strlen30(zName) - 1; 1658 while( zName[i]!='_' ) i--; 1659 zName += i; 1660 } 1661 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 1662 }else{ 1663 sqlite3DebugPrintf("%20s",""); 1664 } 1665 }else{ 1666 char *z; 1667 if( p->u.vtab.idxStr ){ 1668 z = sqlite3_mprintf("(%d,\"%s\",%x)", 1669 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 1670 }else{ 1671 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 1672 } 1673 sqlite3DebugPrintf(" %-19s", z); 1674 sqlite3_free(z); 1675 } 1676 if( p->wsFlags & WHERE_SKIPSCAN ){ 1677 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 1678 }else{ 1679 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); 1680 } 1681 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 1682 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 1683 int i; 1684 for(i=0; i<p->nLTerm; i++){ 1685 whereTermPrint(p->aLTerm[i], i); 1686 } 1687 } 1688 } 1689 #endif 1690 1691 /* 1692 ** Convert bulk memory into a valid WhereLoop that can be passed 1693 ** to whereLoopClear harmlessly. 1694 */ 1695 static void whereLoopInit(WhereLoop *p){ 1696 p->aLTerm = p->aLTermSpace; 1697 p->nLTerm = 0; 1698 p->nLSlot = ArraySize(p->aLTermSpace); 1699 p->wsFlags = 0; 1700 } 1701 1702 /* 1703 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 1704 */ 1705 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 1706 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 1707 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 1708 sqlite3_free(p->u.vtab.idxStr); 1709 p->u.vtab.needFree = 0; 1710 p->u.vtab.idxStr = 0; 1711 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 1712 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 1713 sqlite3DbFree(db, p->u.btree.pIndex); 1714 p->u.btree.pIndex = 0; 1715 } 1716 } 1717 } 1718 1719 /* 1720 ** Deallocate internal memory used by a WhereLoop object 1721 */ 1722 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 1723 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFree(db, p->aLTerm); 1724 whereLoopClearUnion(db, p); 1725 whereLoopInit(p); 1726 } 1727 1728 /* 1729 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 1730 */ 1731 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 1732 WhereTerm **paNew; 1733 if( p->nLSlot>=n ) return SQLITE_OK; 1734 n = (n+7)&~7; 1735 paNew = sqlite3DbMallocRaw(db, sizeof(p->aLTerm[0])*n); 1736 if( paNew==0 ) return SQLITE_NOMEM; 1737 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 1738 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFree(db, p->aLTerm); 1739 p->aLTerm = paNew; 1740 p->nLSlot = n; 1741 return SQLITE_OK; 1742 } 1743 1744 /* 1745 ** Transfer content from the second pLoop into the first. 1746 */ 1747 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 1748 whereLoopClearUnion(db, pTo); 1749 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 1750 memset(&pTo->u, 0, sizeof(pTo->u)); 1751 return SQLITE_NOMEM; 1752 } 1753 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 1754 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 1755 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 1756 pFrom->u.vtab.needFree = 0; 1757 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 1758 pFrom->u.btree.pIndex = 0; 1759 } 1760 return SQLITE_OK; 1761 } 1762 1763 /* 1764 ** Delete a WhereLoop object 1765 */ 1766 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 1767 whereLoopClear(db, p); 1768 sqlite3DbFree(db, p); 1769 } 1770 1771 /* 1772 ** Free a WhereInfo structure 1773 */ 1774 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 1775 if( ALWAYS(pWInfo) ){ 1776 int i; 1777 for(i=0; i<pWInfo->nLevel; i++){ 1778 WhereLevel *pLevel = &pWInfo->a[i]; 1779 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){ 1780 sqlite3DbFree(db, pLevel->u.in.aInLoop); 1781 } 1782 } 1783 sqlite3WhereClauseClear(&pWInfo->sWC); 1784 while( pWInfo->pLoops ){ 1785 WhereLoop *p = pWInfo->pLoops; 1786 pWInfo->pLoops = p->pNextLoop; 1787 whereLoopDelete(db, p); 1788 } 1789 sqlite3DbFree(db, pWInfo); 1790 } 1791 } 1792 1793 /* 1794 ** Return TRUE if all of the following are true: 1795 ** 1796 ** (1) X has the same or lower cost that Y 1797 ** (2) X is a proper subset of Y 1798 ** (3) X skips at least as many columns as Y 1799 ** 1800 ** By "proper subset" we mean that X uses fewer WHERE clause terms 1801 ** than Y and that every WHERE clause term used by X is also used 1802 ** by Y. 1803 ** 1804 ** If X is a proper subset of Y then Y is a better choice and ought 1805 ** to have a lower cost. This routine returns TRUE when that cost 1806 ** relationship is inverted and needs to be adjusted. The third rule 1807 ** was added because if X uses skip-scan less than Y it still might 1808 ** deserve a lower cost even if it is a proper subset of Y. 1809 */ 1810 static int whereLoopCheaperProperSubset( 1811 const WhereLoop *pX, /* First WhereLoop to compare */ 1812 const WhereLoop *pY /* Compare against this WhereLoop */ 1813 ){ 1814 int i, j; 1815 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 1816 return 0; /* X is not a subset of Y */ 1817 } 1818 if( pY->nSkip > pX->nSkip ) return 0; 1819 if( pX->rRun >= pY->rRun ){ 1820 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ 1821 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ 1822 } 1823 for(i=pX->nLTerm-1; i>=0; i--){ 1824 if( pX->aLTerm[i]==0 ) continue; 1825 for(j=pY->nLTerm-1; j>=0; j--){ 1826 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 1827 } 1828 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 1829 } 1830 return 1; /* All conditions meet */ 1831 } 1832 1833 /* 1834 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so 1835 ** that: 1836 ** 1837 ** (1) pTemplate costs less than any other WhereLoops that are a proper 1838 ** subset of pTemplate 1839 ** 1840 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 1841 ** is a proper subset. 1842 ** 1843 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 1844 ** WHERE clause terms than Y and that every WHERE clause term used by X is 1845 ** also used by Y. 1846 */ 1847 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 1848 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 1849 for(; p; p=p->pNextLoop){ 1850 if( p->iTab!=pTemplate->iTab ) continue; 1851 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 1852 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 1853 /* Adjust pTemplate cost downward so that it is cheaper than its 1854 ** subset p. */ 1855 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 1856 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); 1857 pTemplate->rRun = p->rRun; 1858 pTemplate->nOut = p->nOut - 1; 1859 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 1860 /* Adjust pTemplate cost upward so that it is costlier than p since 1861 ** pTemplate is a proper subset of p */ 1862 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 1863 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); 1864 pTemplate->rRun = p->rRun; 1865 pTemplate->nOut = p->nOut + 1; 1866 } 1867 } 1868 } 1869 1870 /* 1871 ** Search the list of WhereLoops in *ppPrev looking for one that can be 1872 ** supplanted by pTemplate. 1873 ** 1874 ** Return NULL if the WhereLoop list contains an entry that can supplant 1875 ** pTemplate, in other words if pTemplate does not belong on the list. 1876 ** 1877 ** If pX is a WhereLoop that pTemplate can supplant, then return the 1878 ** link that points to pX. 1879 ** 1880 ** If pTemplate cannot supplant any existing element of the list but needs 1881 ** to be added to the list, then return a pointer to the tail of the list. 1882 */ 1883 static WhereLoop **whereLoopFindLesser( 1884 WhereLoop **ppPrev, 1885 const WhereLoop *pTemplate 1886 ){ 1887 WhereLoop *p; 1888 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 1889 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 1890 /* If either the iTab or iSortIdx values for two WhereLoop are different 1891 ** then those WhereLoops need to be considered separately. Neither is 1892 ** a candidate to replace the other. */ 1893 continue; 1894 } 1895 /* In the current implementation, the rSetup value is either zero 1896 ** or the cost of building an automatic index (NlogN) and the NlogN 1897 ** is the same for compatible WhereLoops. */ 1898 assert( p->rSetup==0 || pTemplate->rSetup==0 1899 || p->rSetup==pTemplate->rSetup ); 1900 1901 /* whereLoopAddBtree() always generates and inserts the automatic index 1902 ** case first. Hence compatible candidate WhereLoops never have a larger 1903 ** rSetup. Call this SETUP-INVARIANT */ 1904 assert( p->rSetup>=pTemplate->rSetup ); 1905 1906 /* Any loop using an appliation-defined index (or PRIMARY KEY or 1907 ** UNIQUE constraint) with one or more == constraints is better 1908 ** than an automatic index. Unless it is a skip-scan. */ 1909 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 1910 && (pTemplate->nSkip)==0 1911 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 1912 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 1913 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 1914 ){ 1915 break; 1916 } 1917 1918 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 1919 ** discarded. WhereLoop p is better if: 1920 ** (1) p has no more dependencies than pTemplate, and 1921 ** (2) p has an equal or lower cost than pTemplate 1922 */ 1923 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 1924 && p->rSetup<=pTemplate->rSetup /* (2a) */ 1925 && p->rRun<=pTemplate->rRun /* (2b) */ 1926 && p->nOut<=pTemplate->nOut /* (2c) */ 1927 ){ 1928 return 0; /* Discard pTemplate */ 1929 } 1930 1931 /* If pTemplate is always better than p, then cause p to be overwritten 1932 ** with pTemplate. pTemplate is better than p if: 1933 ** (1) pTemplate has no more dependences than p, and 1934 ** (2) pTemplate has an equal or lower cost than p. 1935 */ 1936 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 1937 && p->rRun>=pTemplate->rRun /* (2a) */ 1938 && p->nOut>=pTemplate->nOut /* (2b) */ 1939 ){ 1940 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 1941 break; /* Cause p to be overwritten by pTemplate */ 1942 } 1943 } 1944 return ppPrev; 1945 } 1946 1947 /* 1948 ** Insert or replace a WhereLoop entry using the template supplied. 1949 ** 1950 ** An existing WhereLoop entry might be overwritten if the new template 1951 ** is better and has fewer dependencies. Or the template will be ignored 1952 ** and no insert will occur if an existing WhereLoop is faster and has 1953 ** fewer dependencies than the template. Otherwise a new WhereLoop is 1954 ** added based on the template. 1955 ** 1956 ** If pBuilder->pOrSet is not NULL then we care about only the 1957 ** prerequisites and rRun and nOut costs of the N best loops. That 1958 ** information is gathered in the pBuilder->pOrSet object. This special 1959 ** processing mode is used only for OR clause processing. 1960 ** 1961 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 1962 ** still might overwrite similar loops with the new template if the 1963 ** new template is better. Loops may be overwritten if the following 1964 ** conditions are met: 1965 ** 1966 ** (1) They have the same iTab. 1967 ** (2) They have the same iSortIdx. 1968 ** (3) The template has same or fewer dependencies than the current loop 1969 ** (4) The template has the same or lower cost than the current loop 1970 */ 1971 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 1972 WhereLoop **ppPrev, *p; 1973 WhereInfo *pWInfo = pBuilder->pWInfo; 1974 sqlite3 *db = pWInfo->pParse->db; 1975 1976 /* If pBuilder->pOrSet is defined, then only keep track of the costs 1977 ** and prereqs. 1978 */ 1979 if( pBuilder->pOrSet!=0 ){ 1980 if( pTemplate->nLTerm ){ 1981 #if WHERETRACE_ENABLED 1982 u16 n = pBuilder->pOrSet->n; 1983 int x = 1984 #endif 1985 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 1986 pTemplate->nOut); 1987 #if WHERETRACE_ENABLED /* 0x8 */ 1988 if( sqlite3WhereTrace & 0x8 ){ 1989 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 1990 whereLoopPrint(pTemplate, pBuilder->pWC); 1991 } 1992 #endif 1993 } 1994 return SQLITE_OK; 1995 } 1996 1997 /* Look for an existing WhereLoop to replace with pTemplate 1998 */ 1999 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2000 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2001 2002 if( ppPrev==0 ){ 2003 /* There already exists a WhereLoop on the list that is better 2004 ** than pTemplate, so just ignore pTemplate */ 2005 #if WHERETRACE_ENABLED /* 0x8 */ 2006 if( sqlite3WhereTrace & 0x8 ){ 2007 sqlite3DebugPrintf(" skip: "); 2008 whereLoopPrint(pTemplate, pBuilder->pWC); 2009 } 2010 #endif 2011 return SQLITE_OK; 2012 }else{ 2013 p = *ppPrev; 2014 } 2015 2016 /* If we reach this point it means that either p[] should be overwritten 2017 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2018 ** WhereLoop and insert it. 2019 */ 2020 #if WHERETRACE_ENABLED /* 0x8 */ 2021 if( sqlite3WhereTrace & 0x8 ){ 2022 if( p!=0 ){ 2023 sqlite3DebugPrintf("replace: "); 2024 whereLoopPrint(p, pBuilder->pWC); 2025 } 2026 sqlite3DebugPrintf(" add: "); 2027 whereLoopPrint(pTemplate, pBuilder->pWC); 2028 } 2029 #endif 2030 if( p==0 ){ 2031 /* Allocate a new WhereLoop to add to the end of the list */ 2032 *ppPrev = p = sqlite3DbMallocRaw(db, sizeof(WhereLoop)); 2033 if( p==0 ) return SQLITE_NOMEM; 2034 whereLoopInit(p); 2035 p->pNextLoop = 0; 2036 }else{ 2037 /* We will be overwriting WhereLoop p[]. But before we do, first 2038 ** go through the rest of the list and delete any other entries besides 2039 ** p[] that are also supplated by pTemplate */ 2040 WhereLoop **ppTail = &p->pNextLoop; 2041 WhereLoop *pToDel; 2042 while( *ppTail ){ 2043 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2044 if( ppTail==0 ) break; 2045 pToDel = *ppTail; 2046 if( pToDel==0 ) break; 2047 *ppTail = pToDel->pNextLoop; 2048 #if WHERETRACE_ENABLED /* 0x8 */ 2049 if( sqlite3WhereTrace & 0x8 ){ 2050 sqlite3DebugPrintf(" delete: "); 2051 whereLoopPrint(pToDel, pBuilder->pWC); 2052 } 2053 #endif 2054 whereLoopDelete(db, pToDel); 2055 } 2056 } 2057 whereLoopXfer(db, p, pTemplate); 2058 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2059 Index *pIndex = p->u.btree.pIndex; 2060 if( pIndex && pIndex->tnum==0 ){ 2061 p->u.btree.pIndex = 0; 2062 } 2063 } 2064 return SQLITE_OK; 2065 } 2066 2067 /* 2068 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2069 ** WHERE clause that reference the loop but which are not used by an 2070 ** index. 2071 * 2072 ** For every WHERE clause term that is not used by the index 2073 ** and which has a truth probability assigned by one of the likelihood(), 2074 ** likely(), or unlikely() SQL functions, reduce the estimated number 2075 ** of output rows by the probability specified. 2076 ** 2077 ** TUNING: For every WHERE clause term that is not used by the index 2078 ** and which does not have an assigned truth probability, heuristics 2079 ** described below are used to try to estimate the truth probability. 2080 ** TODO --> Perhaps this is something that could be improved by better 2081 ** table statistics. 2082 ** 2083 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2084 ** value corresponds to -1 in LogEst notation, so this means decrement 2085 ** the WhereLoop.nOut field for every such WHERE clause term. 2086 ** 2087 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2088 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2089 ** final output row estimate is no greater than 1/4 of the total number 2090 ** of rows in the table. In other words, assume that x==EXPR will filter 2091 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2092 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2093 ** on the "x" column and so in that case only cap the output row estimate 2094 ** at 1/2 instead of 1/4. 2095 */ 2096 static void whereLoopOutputAdjust( 2097 WhereClause *pWC, /* The WHERE clause */ 2098 WhereLoop *pLoop, /* The loop to adjust downward */ 2099 LogEst nRow /* Number of rows in the entire table */ 2100 ){ 2101 WhereTerm *pTerm, *pX; 2102 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2103 int i, j, k; 2104 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2105 2106 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2107 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ 2108 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; 2109 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2110 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2111 for(j=pLoop->nLTerm-1; j>=0; j--){ 2112 pX = pLoop->aLTerm[j]; 2113 if( pX==0 ) continue; 2114 if( pX==pTerm ) break; 2115 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2116 } 2117 if( j<0 ){ 2118 if( pTerm->truthProb<=0 ){ 2119 /* If a truth probability is specified using the likelihood() hints, 2120 ** then use the probability provided by the application. */ 2121 pLoop->nOut += pTerm->truthProb; 2122 }else{ 2123 /* In the absence of explicit truth probabilities, use heuristics to 2124 ** guess a reasonable truth probability. */ 2125 pLoop->nOut--; 2126 if( pTerm->eOperator&(WO_EQ|WO_IS) ){ 2127 Expr *pRight = pTerm->pExpr->pRight; 2128 testcase( pTerm->pExpr->op==TK_IS ); 2129 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2130 k = 10; 2131 }else{ 2132 k = 20; 2133 } 2134 if( iReduce<k ) iReduce = k; 2135 } 2136 } 2137 } 2138 } 2139 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce; 2140 } 2141 2142 /* 2143 ** Adjust the cost C by the costMult facter T. This only occurs if 2144 ** compiled with -DSQLITE_ENABLE_COSTMULT 2145 */ 2146 #ifdef SQLITE_ENABLE_COSTMULT 2147 # define ApplyCostMultiplier(C,T) C += T 2148 #else 2149 # define ApplyCostMultiplier(C,T) 2150 #endif 2151 2152 /* 2153 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2154 ** index pIndex. Try to match one more. 2155 ** 2156 ** When this function is called, pBuilder->pNew->nOut contains the 2157 ** number of rows expected to be visited by filtering using the nEq 2158 ** terms only. If it is modified, this value is restored before this 2159 ** function returns. 2160 ** 2161 ** If pProbe->tnum==0, that means pIndex is a fake index used for the 2162 ** INTEGER PRIMARY KEY. 2163 */ 2164 static int whereLoopAddBtreeIndex( 2165 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2166 struct SrcList_item *pSrc, /* FROM clause term being analyzed */ 2167 Index *pProbe, /* An index on pSrc */ 2168 LogEst nInMul /* log(Number of iterations due to IN) */ 2169 ){ 2170 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2171 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2172 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2173 WhereLoop *pNew; /* Template WhereLoop under construction */ 2174 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2175 int opMask; /* Valid operators for constraints */ 2176 WhereScan scan; /* Iterator for WHERE terms */ 2177 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2178 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2179 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2180 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2181 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2182 LogEst saved_nOut; /* Original value of pNew->nOut */ 2183 int rc = SQLITE_OK; /* Return code */ 2184 LogEst rSize; /* Number of rows in the table */ 2185 LogEst rLogSize; /* Logarithm of table size */ 2186 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2187 2188 pNew = pBuilder->pNew; 2189 if( db->mallocFailed ) return SQLITE_NOMEM; 2190 2191 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2192 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2193 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2194 opMask = WO_LT|WO_LE; 2195 }else if( /*pProbe->tnum<=0 ||*/ (pSrc->fg.jointype & JT_LEFT)!=0 ){ 2196 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE; 2197 }else{ 2198 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2199 } 2200 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2201 2202 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2203 2204 saved_nEq = pNew->u.btree.nEq; 2205 saved_nSkip = pNew->nSkip; 2206 saved_nLTerm = pNew->nLTerm; 2207 saved_wsFlags = pNew->wsFlags; 2208 saved_prereq = pNew->prereq; 2209 saved_nOut = pNew->nOut; 2210 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2211 opMask, pProbe); 2212 pNew->rSetup = 0; 2213 rSize = pProbe->aiRowLogEst[0]; 2214 rLogSize = estLog(rSize); 2215 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2216 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2217 LogEst rCostIdx; 2218 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2219 int nIn = 0; 2220 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2221 int nRecValid = pBuilder->nRecValid; 2222 #endif 2223 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2224 && indexColumnNotNull(pProbe, saved_nEq) 2225 ){ 2226 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2227 } 2228 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2229 2230 /* Do not allow the upper bound of a LIKE optimization range constraint 2231 ** to mix with a lower range bound from some other source */ 2232 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2233 2234 pNew->wsFlags = saved_wsFlags; 2235 pNew->u.btree.nEq = saved_nEq; 2236 pNew->nLTerm = saved_nLTerm; 2237 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2238 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2239 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2240 2241 assert( nInMul==0 2242 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2243 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2244 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2245 ); 2246 2247 if( eOp & WO_IN ){ 2248 Expr *pExpr = pTerm->pExpr; 2249 pNew->wsFlags |= WHERE_COLUMN_IN; 2250 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 2251 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2252 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2253 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2254 /* "x IN (value, value, ...)" */ 2255 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2256 } 2257 assert( nIn>0 ); /* RHS always has 2 or more terms... The parser 2258 ** changes "x IN (?)" into "x=?". */ 2259 2260 }else if( eOp & (WO_EQ|WO_IS) ){ 2261 int iCol = pProbe->aiColumn[saved_nEq]; 2262 pNew->wsFlags |= WHERE_COLUMN_EQ; 2263 assert( saved_nEq==pNew->u.btree.nEq ); 2264 if( iCol==XN_ROWID 2265 || (iCol>0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2266 ){ 2267 if( iCol>=0 && pProbe->uniqNotNull==0 ){ 2268 pNew->wsFlags |= WHERE_UNQ_WANTED; 2269 }else{ 2270 pNew->wsFlags |= WHERE_ONEROW; 2271 } 2272 } 2273 }else if( eOp & WO_ISNULL ){ 2274 pNew->wsFlags |= WHERE_COLUMN_NULL; 2275 }else if( eOp & (WO_GT|WO_GE) ){ 2276 testcase( eOp & WO_GT ); 2277 testcase( eOp & WO_GE ); 2278 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2279 pBtm = pTerm; 2280 pTop = 0; 2281 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2282 /* Range contraints that come from the LIKE optimization are 2283 ** always used in pairs. */ 2284 pTop = &pTerm[1]; 2285 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2286 assert( pTop->wtFlags & TERM_LIKEOPT ); 2287 assert( pTop->eOperator==WO_LT ); 2288 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2289 pNew->aLTerm[pNew->nLTerm++] = pTop; 2290 pNew->wsFlags |= WHERE_TOP_LIMIT; 2291 } 2292 }else{ 2293 assert( eOp & (WO_LT|WO_LE) ); 2294 testcase( eOp & WO_LT ); 2295 testcase( eOp & WO_LE ); 2296 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2297 pTop = pTerm; 2298 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2299 pNew->aLTerm[pNew->nLTerm-2] : 0; 2300 } 2301 2302 /* At this point pNew->nOut is set to the number of rows expected to 2303 ** be visited by the index scan before considering term pTerm, or the 2304 ** values of nIn and nInMul. In other words, assuming that all 2305 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2306 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2307 assert( pNew->nOut==saved_nOut ); 2308 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2309 /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4 2310 ** data, using some other estimate. */ 2311 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2312 }else{ 2313 int nEq = ++pNew->u.btree.nEq; 2314 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2315 2316 assert( pNew->nOut==saved_nOut ); 2317 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2318 assert( (eOp & WO_IN) || nIn==0 ); 2319 testcase( eOp & WO_IN ); 2320 pNew->nOut += pTerm->truthProb; 2321 pNew->nOut -= nIn; 2322 }else{ 2323 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2324 tRowcnt nOut = 0; 2325 if( nInMul==0 2326 && pProbe->nSample 2327 && pNew->u.btree.nEq<=pProbe->nSampleCol 2328 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) 2329 ){ 2330 Expr *pExpr = pTerm->pExpr; 2331 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2332 testcase( eOp & WO_EQ ); 2333 testcase( eOp & WO_IS ); 2334 testcase( eOp & WO_ISNULL ); 2335 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2336 }else{ 2337 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2338 } 2339 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2340 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2341 if( nOut ){ 2342 pNew->nOut = sqlite3LogEst(nOut); 2343 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2344 pNew->nOut -= nIn; 2345 } 2346 } 2347 if( nOut==0 ) 2348 #endif 2349 { 2350 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2351 if( eOp & WO_ISNULL ){ 2352 /* TUNING: If there is no likelihood() value, assume that a 2353 ** "col IS NULL" expression matches twice as many rows 2354 ** as (col=?). */ 2355 pNew->nOut += 10; 2356 } 2357 } 2358 } 2359 } 2360 2361 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2362 ** it to pNew->rRun, which is currently set to the cost of the index 2363 ** seek only. Then, if this is a non-covering index, add the cost of 2364 ** visiting the rows in the main table. */ 2365 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2366 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2367 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2368 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2369 } 2370 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2371 2372 nOutUnadjusted = pNew->nOut; 2373 pNew->rRun += nInMul + nIn; 2374 pNew->nOut += nInMul + nIn; 2375 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 2376 rc = whereLoopInsert(pBuilder, pNew); 2377 2378 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2379 pNew->nOut = saved_nOut; 2380 }else{ 2381 pNew->nOut = nOutUnadjusted; 2382 } 2383 2384 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 2385 && pNew->u.btree.nEq<pProbe->nColumn 2386 ){ 2387 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 2388 } 2389 pNew->nOut = saved_nOut; 2390 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2391 pBuilder->nRecValid = nRecValid; 2392 #endif 2393 } 2394 pNew->prereq = saved_prereq; 2395 pNew->u.btree.nEq = saved_nEq; 2396 pNew->nSkip = saved_nSkip; 2397 pNew->wsFlags = saved_wsFlags; 2398 pNew->nOut = saved_nOut; 2399 pNew->nLTerm = saved_nLTerm; 2400 2401 /* Consider using a skip-scan if there are no WHERE clause constraints 2402 ** available for the left-most terms of the index, and if the average 2403 ** number of repeats in the left-most terms is at least 18. 2404 ** 2405 ** The magic number 18 is selected on the basis that scanning 17 rows 2406 ** is almost always quicker than an index seek (even though if the index 2407 ** contains fewer than 2^17 rows we assume otherwise in other parts of 2408 ** the code). And, even if it is not, it should not be too much slower. 2409 ** On the other hand, the extra seeks could end up being significantly 2410 ** more expensive. */ 2411 assert( 42==sqlite3LogEst(18) ); 2412 if( saved_nEq==saved_nSkip 2413 && saved_nEq+1<pProbe->nKeyCol 2414 && pProbe->noSkipScan==0 2415 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 2416 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 2417 ){ 2418 LogEst nIter; 2419 pNew->u.btree.nEq++; 2420 pNew->nSkip++; 2421 pNew->aLTerm[pNew->nLTerm++] = 0; 2422 pNew->wsFlags |= WHERE_SKIPSCAN; 2423 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 2424 pNew->nOut -= nIter; 2425 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 2426 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 2427 nIter += 5; 2428 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 2429 pNew->nOut = saved_nOut; 2430 pNew->u.btree.nEq = saved_nEq; 2431 pNew->nSkip = saved_nSkip; 2432 pNew->wsFlags = saved_wsFlags; 2433 } 2434 2435 return rc; 2436 } 2437 2438 /* 2439 ** Return True if it is possible that pIndex might be useful in 2440 ** implementing the ORDER BY clause in pBuilder. 2441 ** 2442 ** Return False if pBuilder does not contain an ORDER BY clause or 2443 ** if there is no way for pIndex to be useful in implementing that 2444 ** ORDER BY clause. 2445 */ 2446 static int indexMightHelpWithOrderBy( 2447 WhereLoopBuilder *pBuilder, 2448 Index *pIndex, 2449 int iCursor 2450 ){ 2451 ExprList *pOB; 2452 ExprList *aColExpr; 2453 int ii, jj; 2454 2455 if( pIndex->bUnordered ) return 0; 2456 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 2457 for(ii=0; ii<pOB->nExpr; ii++){ 2458 Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr); 2459 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 2460 if( pExpr->iColumn<0 ) return 1; 2461 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2462 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 2463 } 2464 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 2465 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2466 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 2467 if( sqlite3ExprCompare(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 2468 return 1; 2469 } 2470 } 2471 } 2472 } 2473 return 0; 2474 } 2475 2476 /* 2477 ** Return a bitmask where 1s indicate that the corresponding column of 2478 ** the table is used by an index. Only the first 63 columns are considered. 2479 */ 2480 static Bitmask columnsInIndex(Index *pIdx){ 2481 Bitmask m = 0; 2482 int j; 2483 for(j=pIdx->nColumn-1; j>=0; j--){ 2484 int x = pIdx->aiColumn[j]; 2485 if( x>=0 ){ 2486 testcase( x==BMS-1 ); 2487 testcase( x==BMS-2 ); 2488 if( x<BMS-1 ) m |= MASKBIT(x); 2489 } 2490 } 2491 return m; 2492 } 2493 2494 /* Check to see if a partial index with pPartIndexWhere can be used 2495 ** in the current query. Return true if it can be and false if not. 2496 */ 2497 static int whereUsablePartialIndex(int iTab, WhereClause *pWC, Expr *pWhere){ 2498 int i; 2499 WhereTerm *pTerm; 2500 while( pWhere->op==TK_AND ){ 2501 if( !whereUsablePartialIndex(iTab,pWC,pWhere->pLeft) ) return 0; 2502 pWhere = pWhere->pRight; 2503 } 2504 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2505 Expr *pExpr = pTerm->pExpr; 2506 if( sqlite3ExprImpliesExpr(pExpr, pWhere, iTab) 2507 && (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 2508 ){ 2509 return 1; 2510 } 2511 } 2512 return 0; 2513 } 2514 2515 /* 2516 ** Add all WhereLoop objects for a single table of the join where the table 2517 ** is idenfied by pBuilder->pNew->iTab. That table is guaranteed to be 2518 ** a b-tree table, not a virtual table. 2519 ** 2520 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 2521 ** are calculated as follows: 2522 ** 2523 ** For a full scan, assuming the table (or index) contains nRow rows: 2524 ** 2525 ** cost = nRow * 3.0 // full-table scan 2526 ** cost = nRow * K // scan of covering index 2527 ** cost = nRow * (K+3.0) // scan of non-covering index 2528 ** 2529 ** where K is a value between 1.1 and 3.0 set based on the relative 2530 ** estimated average size of the index and table records. 2531 ** 2532 ** For an index scan, where nVisit is the number of index rows visited 2533 ** by the scan, and nSeek is the number of seek operations required on 2534 ** the index b-tree: 2535 ** 2536 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 2537 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 2538 ** 2539 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 2540 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 2541 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 2542 ** 2543 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 2544 ** of uncertainty. For this reason, scoring is designed to pick plans that 2545 ** "do the least harm" if the estimates are inaccurate. For example, a 2546 ** log(nRow) factor is omitted from a non-covering index scan in order to 2547 ** bias the scoring in favor of using an index, since the worst-case 2548 ** performance of using an index is far better than the worst-case performance 2549 ** of a full table scan. 2550 */ 2551 static int whereLoopAddBtree( 2552 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2553 Bitmask mExtra /* Extra prerequesites for using this table */ 2554 ){ 2555 WhereInfo *pWInfo; /* WHERE analysis context */ 2556 Index *pProbe; /* An index we are evaluating */ 2557 Index sPk; /* A fake index object for the primary key */ 2558 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 2559 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 2560 SrcList *pTabList; /* The FROM clause */ 2561 struct SrcList_item *pSrc; /* The FROM clause btree term to add */ 2562 WhereLoop *pNew; /* Template WhereLoop object */ 2563 int rc = SQLITE_OK; /* Return code */ 2564 int iSortIdx = 1; /* Index number */ 2565 int b; /* A boolean value */ 2566 LogEst rSize; /* number of rows in the table */ 2567 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 2568 WhereClause *pWC; /* The parsed WHERE clause */ 2569 Table *pTab; /* Table being queried */ 2570 2571 pNew = pBuilder->pNew; 2572 pWInfo = pBuilder->pWInfo; 2573 pTabList = pWInfo->pTabList; 2574 pSrc = pTabList->a + pNew->iTab; 2575 pTab = pSrc->pTab; 2576 pWC = pBuilder->pWC; 2577 assert( !IsVirtual(pSrc->pTab) ); 2578 2579 if( pSrc->pIBIndex ){ 2580 /* An INDEXED BY clause specifies a particular index to use */ 2581 pProbe = pSrc->pIBIndex; 2582 }else if( !HasRowid(pTab) ){ 2583 pProbe = pTab->pIndex; 2584 }else{ 2585 /* There is no INDEXED BY clause. Create a fake Index object in local 2586 ** variable sPk to represent the rowid primary key index. Make this 2587 ** fake index the first in a chain of Index objects with all of the real 2588 ** indices to follow */ 2589 Index *pFirst; /* First of real indices on the table */ 2590 memset(&sPk, 0, sizeof(Index)); 2591 sPk.nKeyCol = 1; 2592 sPk.nColumn = 1; 2593 sPk.aiColumn = &aiColumnPk; 2594 sPk.aiRowLogEst = aiRowEstPk; 2595 sPk.onError = OE_Replace; 2596 sPk.pTable = pTab; 2597 sPk.szIdxRow = pTab->szTabRow; 2598 aiRowEstPk[0] = pTab->nRowLogEst; 2599 aiRowEstPk[1] = 0; 2600 pFirst = pSrc->pTab->pIndex; 2601 if( pSrc->fg.notIndexed==0 ){ 2602 /* The real indices of the table are only considered if the 2603 ** NOT INDEXED qualifier is omitted from the FROM clause */ 2604 sPk.pNext = pFirst; 2605 } 2606 pProbe = &sPk; 2607 } 2608 rSize = pTab->nRowLogEst; 2609 rLogSize = estLog(rSize); 2610 2611 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 2612 /* Automatic indexes */ 2613 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 2614 && (pWInfo->wctrlFlags & WHERE_NO_AUTOINDEX)==0 2615 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 2616 && pSrc->pIBIndex==0 /* Has no INDEXED BY clause */ 2617 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 2618 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 2619 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 2620 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 2621 ){ 2622 /* Generate auto-index WhereLoops */ 2623 WhereTerm *pTerm; 2624 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 2625 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 2626 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2627 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 2628 pNew->u.btree.nEq = 1; 2629 pNew->nSkip = 0; 2630 pNew->u.btree.pIndex = 0; 2631 pNew->nLTerm = 1; 2632 pNew->aLTerm[0] = pTerm; 2633 /* TUNING: One-time cost for computing the automatic index is 2634 ** estimated to be X*N*log2(N) where N is the number of rows in 2635 ** the table being indexed and where X is 7 (LogEst=28) for normal 2636 ** tables or 1.375 (LogEst=4) for views and subqueries. The value 2637 ** of X is smaller for views and subqueries so that the query planner 2638 ** will be more aggressive about generating automatic indexes for 2639 ** those objects, since there is no opportunity to add schema 2640 ** indexes on subqueries and views. */ 2641 pNew->rSetup = rLogSize + rSize + 4; 2642 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ 2643 pNew->rSetup += 24; 2644 } 2645 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 2646 /* TUNING: Each index lookup yields 20 rows in the table. This 2647 ** is more than the usual guess of 10 rows, since we have no way 2648 ** of knowing how selective the index will ultimately be. It would 2649 ** not be unreasonable to make this value much larger. */ 2650 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 2651 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 2652 pNew->wsFlags = WHERE_AUTO_INDEX; 2653 pNew->prereq = mExtra | pTerm->prereqRight; 2654 rc = whereLoopInsert(pBuilder, pNew); 2655 } 2656 } 2657 } 2658 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 2659 2660 /* Loop over all indices 2661 */ 2662 for(; rc==SQLITE_OK && pProbe; pProbe=pProbe->pNext, iSortIdx++){ 2663 if( pProbe->pPartIdxWhere!=0 2664 && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){ 2665 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 2666 continue; /* Partial index inappropriate for this query */ 2667 } 2668 rSize = pProbe->aiRowLogEst[0]; 2669 pNew->u.btree.nEq = 0; 2670 pNew->nSkip = 0; 2671 pNew->nLTerm = 0; 2672 pNew->iSortIdx = 0; 2673 pNew->rSetup = 0; 2674 pNew->prereq = mExtra; 2675 pNew->nOut = rSize; 2676 pNew->u.btree.pIndex = pProbe; 2677 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 2678 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 2679 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 2680 if( pProbe->tnum<=0 ){ 2681 /* Integer primary key index */ 2682 pNew->wsFlags = WHERE_IPK; 2683 2684 /* Full table scan */ 2685 pNew->iSortIdx = b ? iSortIdx : 0; 2686 /* TUNING: Cost of full table scan is (N*3.0). */ 2687 pNew->rRun = rSize + 16; 2688 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2689 whereLoopOutputAdjust(pWC, pNew, rSize); 2690 rc = whereLoopInsert(pBuilder, pNew); 2691 pNew->nOut = rSize; 2692 if( rc ) break; 2693 }else{ 2694 Bitmask m; 2695 if( pProbe->isCovering ){ 2696 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 2697 m = 0; 2698 }else{ 2699 m = pSrc->colUsed & ~columnsInIndex(pProbe); 2700 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 2701 } 2702 2703 /* Full scan via index */ 2704 if( b 2705 || !HasRowid(pTab) 2706 || ( m==0 2707 && pProbe->bUnordered==0 2708 && (pProbe->szIdxRow<pTab->szTabRow) 2709 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 2710 && sqlite3GlobalConfig.bUseCis 2711 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 2712 ) 2713 ){ 2714 pNew->iSortIdx = b ? iSortIdx : 0; 2715 2716 /* The cost of visiting the index rows is N*K, where K is 2717 ** between 1.1 and 3.0, depending on the relative sizes of the 2718 ** index and table rows. If this is a non-covering index scan, 2719 ** also add the cost of visiting table rows (N*3.0). */ 2720 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 2721 if( m!=0 ){ 2722 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, rSize+16); 2723 } 2724 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2725 whereLoopOutputAdjust(pWC, pNew, rSize); 2726 rc = whereLoopInsert(pBuilder, pNew); 2727 pNew->nOut = rSize; 2728 if( rc ) break; 2729 } 2730 } 2731 2732 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 2733 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2734 sqlite3Stat4ProbeFree(pBuilder->pRec); 2735 pBuilder->nRecValid = 0; 2736 pBuilder->pRec = 0; 2737 #endif 2738 2739 /* If there was an INDEXED BY clause, then only that one index is 2740 ** considered. */ 2741 if( pSrc->pIBIndex ) break; 2742 } 2743 return rc; 2744 } 2745 2746 #ifndef SQLITE_OMIT_VIRTUALTABLE 2747 /* 2748 ** Add all WhereLoop objects for a table of the join identified by 2749 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 2750 ** 2751 ** If there are no LEFT or CROSS JOIN joins in the query, both mExtra and 2752 ** mUnusable are set to 0. Otherwise, mExtra is a mask of all FROM clause 2753 ** entries that occur before the virtual table in the FROM clause and are 2754 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 2755 ** mUnusable mask contains all FROM clause entries that occur after the 2756 ** virtual table and are separated from it by at least one LEFT or 2757 ** CROSS JOIN. 2758 ** 2759 ** For example, if the query were: 2760 ** 2761 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 2762 ** 2763 ** then mExtra corresponds to (t1, t2) and mUnusable to (t5, t6). 2764 ** 2765 ** All the tables in mExtra must be scanned before the current virtual 2766 ** table. So any terms for which all prerequisites are satisfied by 2767 ** mExtra may be specified as "usable" in all calls to xBestIndex. 2768 ** Conversely, all tables in mUnusable must be scanned after the current 2769 ** virtual table, so any terms for which the prerequisites overlap with 2770 ** mUnusable should always be configured as "not-usable" for xBestIndex. 2771 */ 2772 static int whereLoopAddVirtual( 2773 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2774 Bitmask mExtra, /* Tables that must be scanned before this one */ 2775 Bitmask mUnusable /* Tables that must be scanned after this one */ 2776 ){ 2777 WhereInfo *pWInfo; /* WHERE analysis context */ 2778 Parse *pParse; /* The parsing context */ 2779 WhereClause *pWC; /* The WHERE clause */ 2780 struct SrcList_item *pSrc; /* The FROM clause term to search */ 2781 Table *pTab; 2782 sqlite3 *db; 2783 sqlite3_index_info *pIdxInfo; 2784 struct sqlite3_index_constraint *pIdxCons; 2785 struct sqlite3_index_constraint_usage *pUsage; 2786 WhereTerm *pTerm; 2787 int i, j; 2788 int iTerm, mxTerm; 2789 int nConstraint; 2790 int seenIn = 0; /* True if an IN operator is seen */ 2791 int seenVar = 0; /* True if a non-constant constraint is seen */ 2792 int iPhase; /* 0: const w/o IN, 1: const, 2: no IN, 2: IN */ 2793 WhereLoop *pNew; 2794 int rc = SQLITE_OK; 2795 2796 assert( (mExtra & mUnusable)==0 ); 2797 pWInfo = pBuilder->pWInfo; 2798 pParse = pWInfo->pParse; 2799 db = pParse->db; 2800 pWC = pBuilder->pWC; 2801 pNew = pBuilder->pNew; 2802 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 2803 pTab = pSrc->pTab; 2804 assert( IsVirtual(pTab) ); 2805 pIdxInfo = allocateIndexInfo(pParse, pWC, mUnusable, pSrc,pBuilder->pOrderBy); 2806 if( pIdxInfo==0 ) return SQLITE_NOMEM; 2807 pNew->prereq = 0; 2808 pNew->rSetup = 0; 2809 pNew->wsFlags = WHERE_VIRTUALTABLE; 2810 pNew->nLTerm = 0; 2811 pNew->u.vtab.needFree = 0; 2812 pUsage = pIdxInfo->aConstraintUsage; 2813 nConstraint = pIdxInfo->nConstraint; 2814 if( whereLoopResize(db, pNew, nConstraint) ){ 2815 sqlite3DbFree(db, pIdxInfo); 2816 return SQLITE_NOMEM; 2817 } 2818 2819 for(iPhase=0; iPhase<=3; iPhase++){ 2820 if( !seenIn && (iPhase&1)!=0 ){ 2821 iPhase++; 2822 if( iPhase>3 ) break; 2823 } 2824 if( !seenVar && iPhase>1 ) break; 2825 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 2826 for(i=0; i<pIdxInfo->nConstraint; i++, pIdxCons++){ 2827 j = pIdxCons->iTermOffset; 2828 pTerm = &pWC->a[j]; 2829 switch( iPhase ){ 2830 case 0: /* Constants without IN operator */ 2831 pIdxCons->usable = 0; 2832 if( (pTerm->eOperator & WO_IN)!=0 ){ 2833 seenIn = 1; 2834 } 2835 if( (pTerm->prereqRight & ~mExtra)!=0 ){ 2836 seenVar = 1; 2837 }else if( (pTerm->eOperator & WO_IN)==0 ){ 2838 pIdxCons->usable = 1; 2839 } 2840 break; 2841 case 1: /* Constants with IN operators */ 2842 assert( seenIn ); 2843 pIdxCons->usable = (pTerm->prereqRight & ~mExtra)==0; 2844 break; 2845 case 2: /* Variables without IN */ 2846 assert( seenVar ); 2847 pIdxCons->usable = (pTerm->eOperator & WO_IN)==0; 2848 break; 2849 default: /* Variables with IN */ 2850 assert( seenVar && seenIn ); 2851 pIdxCons->usable = 1; 2852 break; 2853 } 2854 } 2855 memset(pUsage, 0, sizeof(pUsage[0])*pIdxInfo->nConstraint); 2856 if( pIdxInfo->needToFreeIdxStr ) sqlite3_free(pIdxInfo->idxStr); 2857 pIdxInfo->idxStr = 0; 2858 pIdxInfo->idxNum = 0; 2859 pIdxInfo->needToFreeIdxStr = 0; 2860 pIdxInfo->orderByConsumed = 0; 2861 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 2862 pIdxInfo->estimatedRows = 25; 2863 pIdxInfo->idxFlags = 0; 2864 rc = vtabBestIndex(pParse, pTab, pIdxInfo); 2865 if( rc ) goto whereLoopAddVtab_exit; 2866 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 2867 pNew->prereq = mExtra; 2868 mxTerm = -1; 2869 assert( pNew->nLSlot>=nConstraint ); 2870 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 2871 pNew->u.vtab.omitMask = 0; 2872 for(i=0; i<nConstraint; i++, pIdxCons++){ 2873 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 2874 j = pIdxCons->iTermOffset; 2875 if( iTerm>=nConstraint 2876 || j<0 2877 || j>=pWC->nTerm 2878 || pNew->aLTerm[iTerm]!=0 2879 ){ 2880 rc = SQLITE_ERROR; 2881 sqlite3ErrorMsg(pParse, "%s.xBestIndex() malfunction", pTab->zName); 2882 goto whereLoopAddVtab_exit; 2883 } 2884 testcase( iTerm==nConstraint-1 ); 2885 testcase( j==0 ); 2886 testcase( j==pWC->nTerm-1 ); 2887 pTerm = &pWC->a[j]; 2888 pNew->prereq |= pTerm->prereqRight; 2889 assert( iTerm<pNew->nLSlot ); 2890 pNew->aLTerm[iTerm] = pTerm; 2891 if( iTerm>mxTerm ) mxTerm = iTerm; 2892 testcase( iTerm==15 ); 2893 testcase( iTerm==16 ); 2894 if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<<iTerm; 2895 if( (pTerm->eOperator & WO_IN)!=0 ){ 2896 if( pUsage[i].omit==0 ){ 2897 /* Do not attempt to use an IN constraint if the virtual table 2898 ** says that the equivalent EQ constraint cannot be safely omitted. 2899 ** If we do attempt to use such a constraint, some rows might be 2900 ** repeated in the output. */ 2901 break; 2902 } 2903 /* A virtual table that is constrained by an IN clause may not 2904 ** consume the ORDER BY clause because (1) the order of IN terms 2905 ** is not necessarily related to the order of output terms and 2906 ** (2) Multiple outputs from a single IN value will not merge 2907 ** together. */ 2908 pIdxInfo->orderByConsumed = 0; 2909 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 2910 } 2911 } 2912 } 2913 if( i>=nConstraint ){ 2914 pNew->nLTerm = mxTerm+1; 2915 assert( pNew->nLTerm<=pNew->nLSlot ); 2916 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 2917 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 2918 pIdxInfo->needToFreeIdxStr = 0; 2919 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 2920 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 2921 pIdxInfo->nOrderBy : 0); 2922 pNew->rSetup = 0; 2923 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 2924 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 2925 2926 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 2927 ** that the scan will visit at most one row. Clear it otherwise. */ 2928 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 2929 pNew->wsFlags |= WHERE_ONEROW; 2930 }else{ 2931 pNew->wsFlags &= ~WHERE_ONEROW; 2932 } 2933 whereLoopInsert(pBuilder, pNew); 2934 if( pNew->u.vtab.needFree ){ 2935 sqlite3_free(pNew->u.vtab.idxStr); 2936 pNew->u.vtab.needFree = 0; 2937 } 2938 } 2939 } 2940 2941 whereLoopAddVtab_exit: 2942 if( pIdxInfo->needToFreeIdxStr ) sqlite3_free(pIdxInfo->idxStr); 2943 sqlite3DbFree(db, pIdxInfo); 2944 return rc; 2945 } 2946 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 2947 2948 /* 2949 ** Add WhereLoop entries to handle OR terms. This works for either 2950 ** btrees or virtual tables. 2951 */ 2952 static int whereLoopAddOr( 2953 WhereLoopBuilder *pBuilder, 2954 Bitmask mExtra, 2955 Bitmask mUnusable 2956 ){ 2957 WhereInfo *pWInfo = pBuilder->pWInfo; 2958 WhereClause *pWC; 2959 WhereLoop *pNew; 2960 WhereTerm *pTerm, *pWCEnd; 2961 int rc = SQLITE_OK; 2962 int iCur; 2963 WhereClause tempWC; 2964 WhereLoopBuilder sSubBuild; 2965 WhereOrSet sSum, sCur; 2966 struct SrcList_item *pItem; 2967 2968 pWC = pBuilder->pWC; 2969 pWCEnd = pWC->a + pWC->nTerm; 2970 pNew = pBuilder->pNew; 2971 memset(&sSum, 0, sizeof(sSum)); 2972 pItem = pWInfo->pTabList->a + pNew->iTab; 2973 iCur = pItem->iCursor; 2974 2975 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 2976 if( (pTerm->eOperator & WO_OR)!=0 2977 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 2978 ){ 2979 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 2980 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 2981 WhereTerm *pOrTerm; 2982 int once = 1; 2983 int i, j; 2984 2985 sSubBuild = *pBuilder; 2986 sSubBuild.pOrderBy = 0; 2987 sSubBuild.pOrSet = &sCur; 2988 2989 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 2990 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 2991 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 2992 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 2993 }else if( pOrTerm->leftCursor==iCur ){ 2994 tempWC.pWInfo = pWC->pWInfo; 2995 tempWC.pOuter = pWC; 2996 tempWC.op = TK_AND; 2997 tempWC.nTerm = 1; 2998 tempWC.a = pOrTerm; 2999 sSubBuild.pWC = &tempWC; 3000 }else{ 3001 continue; 3002 } 3003 sCur.n = 0; 3004 #ifdef WHERETRACE_ENABLED 3005 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3006 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3007 if( sqlite3WhereTrace & 0x400 ){ 3008 for(i=0; i<sSubBuild.pWC->nTerm; i++){ 3009 whereTermPrint(&sSubBuild.pWC->a[i], i); 3010 } 3011 } 3012 #endif 3013 #ifndef SQLITE_OMIT_VIRTUALTABLE 3014 if( IsVirtual(pItem->pTab) ){ 3015 rc = whereLoopAddVirtual(&sSubBuild, mExtra, mUnusable); 3016 }else 3017 #endif 3018 { 3019 rc = whereLoopAddBtree(&sSubBuild, mExtra); 3020 } 3021 if( rc==SQLITE_OK ){ 3022 rc = whereLoopAddOr(&sSubBuild, mExtra, mUnusable); 3023 } 3024 assert( rc==SQLITE_OK || sCur.n==0 ); 3025 if( sCur.n==0 ){ 3026 sSum.n = 0; 3027 break; 3028 }else if( once ){ 3029 whereOrMove(&sSum, &sCur); 3030 once = 0; 3031 }else{ 3032 WhereOrSet sPrev; 3033 whereOrMove(&sPrev, &sSum); 3034 sSum.n = 0; 3035 for(i=0; i<sPrev.n; i++){ 3036 for(j=0; j<sCur.n; j++){ 3037 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3038 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3039 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3040 } 3041 } 3042 } 3043 } 3044 pNew->nLTerm = 1; 3045 pNew->aLTerm[0] = pTerm; 3046 pNew->wsFlags = WHERE_MULTI_OR; 3047 pNew->rSetup = 0; 3048 pNew->iSortIdx = 0; 3049 memset(&pNew->u, 0, sizeof(pNew->u)); 3050 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3051 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3052 ** of all sub-scans required by the OR-scan. However, due to rounding 3053 ** errors, it may be that the cost of the OR-scan is equal to its 3054 ** most expensive sub-scan. Add the smallest possible penalty 3055 ** (equivalent to multiplying the cost by 1.07) to ensure that 3056 ** this does not happen. Otherwise, for WHERE clauses such as the 3057 ** following where there is an index on "y": 3058 ** 3059 ** WHERE likelihood(x=?, 0.99) OR y=? 3060 ** 3061 ** the planner may elect to "OR" together a full-table scan and an 3062 ** index lookup. And other similarly odd results. */ 3063 pNew->rRun = sSum.a[i].rRun + 1; 3064 pNew->nOut = sSum.a[i].nOut; 3065 pNew->prereq = sSum.a[i].prereq; 3066 rc = whereLoopInsert(pBuilder, pNew); 3067 } 3068 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3069 } 3070 } 3071 return rc; 3072 } 3073 3074 /* 3075 ** Add all WhereLoop objects for all tables 3076 */ 3077 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3078 WhereInfo *pWInfo = pBuilder->pWInfo; 3079 Bitmask mExtra = 0; 3080 Bitmask mPrior = 0; 3081 int iTab; 3082 SrcList *pTabList = pWInfo->pTabList; 3083 struct SrcList_item *pItem; 3084 struct SrcList_item *pEnd = &pTabList->a[pWInfo->nLevel]; 3085 sqlite3 *db = pWInfo->pParse->db; 3086 int rc = SQLITE_OK; 3087 WhereLoop *pNew; 3088 u8 priorJointype = 0; 3089 3090 /* Loop over the tables in the join, from left to right */ 3091 pNew = pBuilder->pNew; 3092 whereLoopInit(pNew); 3093 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3094 Bitmask mUnusable = 0; 3095 pNew->iTab = iTab; 3096 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3097 if( ((pItem->fg.jointype|priorJointype) & (JT_LEFT|JT_CROSS))!=0 ){ 3098 /* This condition is true when pItem is the FROM clause term on the 3099 ** right-hand-side of a LEFT or CROSS JOIN. */ 3100 mExtra = mPrior; 3101 } 3102 priorJointype = pItem->fg.jointype; 3103 if( IsVirtual(pItem->pTab) ){ 3104 struct SrcList_item *p; 3105 for(p=&pItem[1]; p<pEnd; p++){ 3106 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3107 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3108 } 3109 } 3110 rc = whereLoopAddVirtual(pBuilder, mExtra, mUnusable); 3111 }else{ 3112 rc = whereLoopAddBtree(pBuilder, mExtra); 3113 } 3114 if( rc==SQLITE_OK ){ 3115 rc = whereLoopAddOr(pBuilder, mExtra, mUnusable); 3116 } 3117 mPrior |= pNew->maskSelf; 3118 if( rc || db->mallocFailed ) break; 3119 } 3120 3121 whereLoopClear(db, pNew); 3122 return rc; 3123 } 3124 3125 /* 3126 ** Examine a WherePath (with the addition of the extra WhereLoop of the 5th 3127 ** parameters) to see if it outputs rows in the requested ORDER BY 3128 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3129 ** 3130 ** N>0: N terms of the ORDER BY clause are satisfied 3131 ** N==0: No terms of the ORDER BY clause are satisfied 3132 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3133 ** 3134 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3135 ** strict. With GROUP BY and DISTINCT the only requirement is that 3136 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3137 ** and DISTINCT do not require rows to appear in any particular order as long 3138 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3139 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3140 ** pOrderBy terms must be matched in strict left-to-right order. 3141 */ 3142 static i8 wherePathSatisfiesOrderBy( 3143 WhereInfo *pWInfo, /* The WHERE clause */ 3144 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3145 WherePath *pPath, /* The WherePath to check */ 3146 u16 wctrlFlags, /* Might contain WHERE_GROUPBY or WHERE_DISTINCTBY */ 3147 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3148 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3149 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3150 ){ 3151 u8 revSet; /* True if rev is known */ 3152 u8 rev; /* Composite sort order */ 3153 u8 revIdx; /* Index sort order */ 3154 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 3155 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 3156 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 3157 u16 nKeyCol; /* Number of key columns in pIndex */ 3158 u16 nColumn; /* Total number of ordered columns in the index */ 3159 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 3160 int iLoop; /* Index of WhereLoop in pPath being processed */ 3161 int i, j; /* Loop counters */ 3162 int iCur; /* Cursor number for current WhereLoop */ 3163 int iColumn; /* A column number within table iCur */ 3164 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 3165 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3166 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 3167 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 3168 Index *pIndex; /* The index associated with pLoop */ 3169 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 3170 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 3171 Bitmask obDone; /* Mask of all ORDER BY terms */ 3172 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 3173 Bitmask ready; /* Mask of inner loops */ 3174 3175 /* 3176 ** We say the WhereLoop is "one-row" if it generates no more than one 3177 ** row of output. A WhereLoop is one-row if all of the following are true: 3178 ** (a) All index columns match with WHERE_COLUMN_EQ. 3179 ** (b) The index is unique 3180 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 3181 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 3182 ** 3183 ** We say the WhereLoop is "order-distinct" if the set of columns from 3184 ** that WhereLoop that are in the ORDER BY clause are different for every 3185 ** row of the WhereLoop. Every one-row WhereLoop is automatically 3186 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 3187 ** is not order-distinct. To be order-distinct is not quite the same as being 3188 ** UNIQUE since a UNIQUE column or index can have multiple rows that 3189 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 3190 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 3191 ** 3192 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 3193 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 3194 ** automatically order-distinct. 3195 */ 3196 3197 assert( pOrderBy!=0 ); 3198 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 3199 3200 nOrderBy = pOrderBy->nExpr; 3201 testcase( nOrderBy==BMS-1 ); 3202 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 3203 isOrderDistinct = 1; 3204 obDone = MASKBIT(nOrderBy)-1; 3205 orderDistinctMask = 0; 3206 ready = 0; 3207 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 3208 if( iLoop>0 ) ready |= pLoop->maskSelf; 3209 pLoop = iLoop<nLoop ? pPath->aLoop[iLoop] : pLast; 3210 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 3211 if( pLoop->u.vtab.isOrdered ) obSat = obDone; 3212 break; 3213 } 3214 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 3215 3216 /* Mark off any ORDER BY term X that is a column in the table of 3217 ** the current loop for which there is term in the WHERE 3218 ** clause of the form X IS NULL or X=? that reference only outer 3219 ** loops. 3220 */ 3221 for(i=0; i<nOrderBy; i++){ 3222 if( MASKBIT(i) & obSat ) continue; 3223 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3224 if( pOBExpr->op!=TK_COLUMN ) continue; 3225 if( pOBExpr->iTable!=iCur ) continue; 3226 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 3227 ~ready, WO_EQ|WO_ISNULL|WO_IS, 0); 3228 if( pTerm==0 ) continue; 3229 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 3230 const char *z1, *z2; 3231 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3232 if( !pColl ) pColl = db->pDfltColl; 3233 z1 = pColl->zName; 3234 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pTerm->pExpr); 3235 if( !pColl ) pColl = db->pDfltColl; 3236 z2 = pColl->zName; 3237 if( sqlite3StrICmp(z1, z2)!=0 ) continue; 3238 testcase( pTerm->pExpr->op==TK_IS ); 3239 } 3240 obSat |= MASKBIT(i); 3241 } 3242 3243 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 3244 if( pLoop->wsFlags & WHERE_IPK ){ 3245 pIndex = 0; 3246 nKeyCol = 0; 3247 nColumn = 1; 3248 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 3249 return 0; 3250 }else{ 3251 nKeyCol = pIndex->nKeyCol; 3252 nColumn = pIndex->nColumn; 3253 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 3254 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 3255 || !HasRowid(pIndex->pTable)); 3256 isOrderDistinct = IsUniqueIndex(pIndex); 3257 } 3258 3259 /* Loop through all columns of the index and deal with the ones 3260 ** that are not constrained by == or IN. 3261 */ 3262 rev = revSet = 0; 3263 distinctColumns = 0; 3264 for(j=0; j<nColumn; j++){ 3265 u8 bOnce; /* True to run the ORDER BY search loop */ 3266 3267 /* Skip over == and IS NULL terms */ 3268 if( j<pLoop->u.btree.nEq 3269 && pLoop->nSkip==0 3270 && ((i = pLoop->aLTerm[j]->eOperator) & (WO_EQ|WO_ISNULL|WO_IS))!=0 3271 ){ 3272 if( i & WO_ISNULL ){ 3273 testcase( isOrderDistinct ); 3274 isOrderDistinct = 0; 3275 } 3276 continue; 3277 } 3278 3279 /* Get the column number in the table (iColumn) and sort order 3280 ** (revIdx) for the j-th column of the index. 3281 */ 3282 if( pIndex ){ 3283 iColumn = pIndex->aiColumn[j]; 3284 revIdx = pIndex->aSortOrder[j]; 3285 if( iColumn==pIndex->pTable->iPKey ) iColumn = -1; 3286 }else{ 3287 iColumn = XN_ROWID; 3288 revIdx = 0; 3289 } 3290 3291 /* An unconstrained column that might be NULL means that this 3292 ** WhereLoop is not well-ordered 3293 */ 3294 if( isOrderDistinct 3295 && iColumn>=0 3296 && j>=pLoop->u.btree.nEq 3297 && pIndex->pTable->aCol[iColumn].notNull==0 3298 ){ 3299 isOrderDistinct = 0; 3300 } 3301 3302 /* Find the ORDER BY term that corresponds to the j-th column 3303 ** of the index and mark that ORDER BY term off 3304 */ 3305 bOnce = 1; 3306 isMatch = 0; 3307 for(i=0; bOnce && i<nOrderBy; i++){ 3308 if( MASKBIT(i) & obSat ) continue; 3309 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3310 testcase( wctrlFlags & WHERE_GROUPBY ); 3311 testcase( wctrlFlags & WHERE_DISTINCTBY ); 3312 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 3313 if( iColumn>=(-1) ){ 3314 if( pOBExpr->op!=TK_COLUMN ) continue; 3315 if( pOBExpr->iTable!=iCur ) continue; 3316 if( pOBExpr->iColumn!=iColumn ) continue; 3317 }else{ 3318 if( sqlite3ExprCompare(pOBExpr,pIndex->aColExpr->a[j].pExpr,iCur) ){ 3319 continue; 3320 } 3321 } 3322 if( iColumn>=0 ){ 3323 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3324 if( !pColl ) pColl = db->pDfltColl; 3325 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 3326 } 3327 isMatch = 1; 3328 break; 3329 } 3330 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 3331 /* Make sure the sort order is compatible in an ORDER BY clause. 3332 ** Sort order is irrelevant for a GROUP BY clause. */ 3333 if( revSet ){ 3334 if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0; 3335 }else{ 3336 rev = revIdx ^ pOrderBy->a[i].sortOrder; 3337 if( rev ) *pRevMask |= MASKBIT(iLoop); 3338 revSet = 1; 3339 } 3340 } 3341 if( isMatch ){ 3342 if( iColumn<0 ){ 3343 testcase( distinctColumns==0 ); 3344 distinctColumns = 1; 3345 } 3346 obSat |= MASKBIT(i); 3347 }else{ 3348 /* No match found */ 3349 if( j==0 || j<nKeyCol ){ 3350 testcase( isOrderDistinct!=0 ); 3351 isOrderDistinct = 0; 3352 } 3353 break; 3354 } 3355 } /* end Loop over all index columns */ 3356 if( distinctColumns ){ 3357 testcase( isOrderDistinct==0 ); 3358 isOrderDistinct = 1; 3359 } 3360 } /* end-if not one-row */ 3361 3362 /* Mark off any other ORDER BY terms that reference pLoop */ 3363 if( isOrderDistinct ){ 3364 orderDistinctMask |= pLoop->maskSelf; 3365 for(i=0; i<nOrderBy; i++){ 3366 Expr *p; 3367 Bitmask mTerm; 3368 if( MASKBIT(i) & obSat ) continue; 3369 p = pOrderBy->a[i].pExpr; 3370 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 3371 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 3372 if( (mTerm&~orderDistinctMask)==0 ){ 3373 obSat |= MASKBIT(i); 3374 } 3375 } 3376 } 3377 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 3378 if( obSat==obDone ) return (i8)nOrderBy; 3379 if( !isOrderDistinct ){ 3380 for(i=nOrderBy-1; i>0; i--){ 3381 Bitmask m = MASKBIT(i) - 1; 3382 if( (obSat&m)==m ) return i; 3383 } 3384 return 0; 3385 } 3386 return -1; 3387 } 3388 3389 3390 /* 3391 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 3392 ** the planner assumes that the specified pOrderBy list is actually a GROUP 3393 ** BY clause - and so any order that groups rows as required satisfies the 3394 ** request. 3395 ** 3396 ** Normally, in this case it is not possible for the caller to determine 3397 ** whether or not the rows are really being delivered in sorted order, or 3398 ** just in some other order that provides the required grouping. However, 3399 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 3400 ** this function may be called on the returned WhereInfo object. It returns 3401 ** true if the rows really will be sorted in the specified order, or false 3402 ** otherwise. 3403 ** 3404 ** For example, assuming: 3405 ** 3406 ** CREATE INDEX i1 ON t1(x, Y); 3407 ** 3408 ** then 3409 ** 3410 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 3411 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 3412 */ 3413 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 3414 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 3415 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 3416 return pWInfo->sorted; 3417 } 3418 3419 #ifdef WHERETRACE_ENABLED 3420 /* For debugging use only: */ 3421 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 3422 static char zName[65]; 3423 int i; 3424 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 3425 if( pLast ) zName[i++] = pLast->cId; 3426 zName[i] = 0; 3427 return zName; 3428 } 3429 #endif 3430 3431 /* 3432 ** Return the cost of sorting nRow rows, assuming that the keys have 3433 ** nOrderby columns and that the first nSorted columns are already in 3434 ** order. 3435 */ 3436 static LogEst whereSortingCost( 3437 WhereInfo *pWInfo, 3438 LogEst nRow, 3439 int nOrderBy, 3440 int nSorted 3441 ){ 3442 /* TUNING: Estimated cost of a full external sort, where N is 3443 ** the number of rows to sort is: 3444 ** 3445 ** cost = (3.0 * N * log(N)). 3446 ** 3447 ** Or, if the order-by clause has X terms but only the last Y 3448 ** terms are out of order, then block-sorting will reduce the 3449 ** sorting cost to: 3450 ** 3451 ** cost = (3.0 * N * log(N)) * (Y/X) 3452 ** 3453 ** The (Y/X) term is implemented using stack variable rScale 3454 ** below. */ 3455 LogEst rScale, rSortCost; 3456 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 3457 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 3458 rSortCost = nRow + estLog(nRow) + rScale + 16; 3459 3460 /* TUNING: The cost of implementing DISTINCT using a B-TREE is 3461 ** similar but with a larger constant of proportionality. 3462 ** Multiply by an additional factor of 3.0. */ 3463 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 3464 rSortCost += 16; 3465 } 3466 3467 return rSortCost; 3468 } 3469 3470 /* 3471 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 3472 ** attempts to find the lowest cost path that visits each WhereLoop 3473 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 3474 ** 3475 ** Assume that the total number of output rows that will need to be sorted 3476 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 3477 ** costs if nRowEst==0. 3478 ** 3479 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 3480 ** error occurs. 3481 */ 3482 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 3483 int mxChoice; /* Maximum number of simultaneous paths tracked */ 3484 int nLoop; /* Number of terms in the join */ 3485 Parse *pParse; /* Parsing context */ 3486 sqlite3 *db; /* The database connection */ 3487 int iLoop; /* Loop counter over the terms of the join */ 3488 int ii, jj; /* Loop counters */ 3489 int mxI = 0; /* Index of next entry to replace */ 3490 int nOrderBy; /* Number of ORDER BY clause terms */ 3491 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 3492 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 3493 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 3494 WherePath *aFrom; /* All nFrom paths at the previous level */ 3495 WherePath *aTo; /* The nTo best paths at the current level */ 3496 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 3497 WherePath *pTo; /* An element of aTo[] that we are working on */ 3498 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 3499 WhereLoop **pX; /* Used to divy up the pSpace memory */ 3500 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 3501 char *pSpace; /* Temporary memory used by this routine */ 3502 int nSpace; /* Bytes of space allocated at pSpace */ 3503 3504 pParse = pWInfo->pParse; 3505 db = pParse->db; 3506 nLoop = pWInfo->nLevel; 3507 /* TUNING: For simple queries, only the best path is tracked. 3508 ** For 2-way joins, the 5 best paths are followed. 3509 ** For joins of 3 or more tables, track the 10 best paths */ 3510 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 3511 assert( nLoop<=pWInfo->pTabList->nSrc ); 3512 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 3513 3514 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 3515 ** case the purpose of this call is to estimate the number of rows returned 3516 ** by the overall query. Once this estimate has been obtained, the caller 3517 ** will invoke this function a second time, passing the estimate as the 3518 ** nRowEst parameter. */ 3519 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 3520 nOrderBy = 0; 3521 }else{ 3522 nOrderBy = pWInfo->pOrderBy->nExpr; 3523 } 3524 3525 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 3526 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 3527 nSpace += sizeof(LogEst) * nOrderBy; 3528 pSpace = sqlite3DbMallocRaw(db, nSpace); 3529 if( pSpace==0 ) return SQLITE_NOMEM; 3530 aTo = (WherePath*)pSpace; 3531 aFrom = aTo+mxChoice; 3532 memset(aFrom, 0, sizeof(aFrom[0])); 3533 pX = (WhereLoop**)(aFrom+mxChoice); 3534 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 3535 pFrom->aLoop = pX; 3536 } 3537 if( nOrderBy ){ 3538 /* If there is an ORDER BY clause and it is not being ignored, set up 3539 ** space for the aSortCost[] array. Each element of the aSortCost array 3540 ** is either zero - meaning it has not yet been initialized - or the 3541 ** cost of sorting nRowEst rows of data where the first X terms of 3542 ** the ORDER BY clause are already in order, where X is the array 3543 ** index. */ 3544 aSortCost = (LogEst*)pX; 3545 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 3546 } 3547 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 3548 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 3549 3550 /* Seed the search with a single WherePath containing zero WhereLoops. 3551 ** 3552 ** TUNING: Do not let the number of iterations go above 28. If the cost 3553 ** of computing an automatic index is not paid back within the first 28 3554 ** rows, then do not use the automatic index. */ 3555 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 3556 nFrom = 1; 3557 assert( aFrom[0].isOrdered==0 ); 3558 if( nOrderBy ){ 3559 /* If nLoop is zero, then there are no FROM terms in the query. Since 3560 ** in this case the query may return a maximum of one row, the results 3561 ** are already in the requested order. Set isOrdered to nOrderBy to 3562 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 3563 ** -1, indicating that the result set may or may not be ordered, 3564 ** depending on the loops added to the current plan. */ 3565 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 3566 } 3567 3568 /* Compute successively longer WherePaths using the previous generation 3569 ** of WherePaths as the basis for the next. Keep track of the mxChoice 3570 ** best paths at each generation */ 3571 for(iLoop=0; iLoop<nLoop; iLoop++){ 3572 nTo = 0; 3573 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 3574 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 3575 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 3576 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 3577 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 3578 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 3579 Bitmask maskNew; /* Mask of src visited by (..) */ 3580 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 3581 3582 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 3583 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 3584 /* At this point, pWLoop is a candidate to be the next loop. 3585 ** Compute its cost */ 3586 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 3587 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 3588 nOut = pFrom->nRow + pWLoop->nOut; 3589 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 3590 if( isOrdered<0 ){ 3591 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 3592 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 3593 iLoop, pWLoop, &revMask); 3594 }else{ 3595 revMask = pFrom->revLoop; 3596 } 3597 if( isOrdered>=0 && isOrdered<nOrderBy ){ 3598 if( aSortCost[isOrdered]==0 ){ 3599 aSortCost[isOrdered] = whereSortingCost( 3600 pWInfo, nRowEst, nOrderBy, isOrdered 3601 ); 3602 } 3603 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]); 3604 3605 WHERETRACE(0x002, 3606 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 3607 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 3608 rUnsorted, rCost)); 3609 }else{ 3610 rCost = rUnsorted; 3611 } 3612 3613 /* Check to see if pWLoop should be added to the set of 3614 ** mxChoice best-so-far paths. 3615 ** 3616 ** First look for an existing path among best-so-far paths 3617 ** that covers the same set of loops and has the same isOrdered 3618 ** setting as the current path candidate. 3619 ** 3620 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 3621 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 3622 ** of legal values for isOrdered, -1..64. 3623 */ 3624 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 3625 if( pTo->maskLoop==maskNew 3626 && ((pTo->isOrdered^isOrdered)&0x80)==0 3627 ){ 3628 testcase( jj==nTo-1 ); 3629 break; 3630 } 3631 } 3632 if( jj>=nTo ){ 3633 /* None of the existing best-so-far paths match the candidate. */ 3634 if( nTo>=mxChoice 3635 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 3636 ){ 3637 /* The current candidate is no better than any of the mxChoice 3638 ** paths currently in the best-so-far buffer. So discard 3639 ** this candidate as not viable. */ 3640 #ifdef WHERETRACE_ENABLED /* 0x4 */ 3641 if( sqlite3WhereTrace&0x4 ){ 3642 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d order=%c\n", 3643 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, 3644 isOrdered>=0 ? isOrdered+'0' : '?'); 3645 } 3646 #endif 3647 continue; 3648 } 3649 /* If we reach this points it means that the new candidate path 3650 ** needs to be added to the set of best-so-far paths. */ 3651 if( nTo<mxChoice ){ 3652 /* Increase the size of the aTo set by one */ 3653 jj = nTo++; 3654 }else{ 3655 /* New path replaces the prior worst to keep count below mxChoice */ 3656 jj = mxI; 3657 } 3658 pTo = &aTo[jj]; 3659 #ifdef WHERETRACE_ENABLED /* 0x4 */ 3660 if( sqlite3WhereTrace&0x4 ){ 3661 sqlite3DebugPrintf("New %s cost=%-3d,%3d order=%c\n", 3662 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, 3663 isOrdered>=0 ? isOrdered+'0' : '?'); 3664 } 3665 #endif 3666 }else{ 3667 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 3668 ** same set of loops and has the sam isOrdered setting as the 3669 ** candidate path. Check to see if the candidate should replace 3670 ** pTo or if the candidate should be skipped */ 3671 if( pTo->rCost<rCost || (pTo->rCost==rCost && pTo->nRow<=nOut) ){ 3672 #ifdef WHERETRACE_ENABLED /* 0x4 */ 3673 if( sqlite3WhereTrace&0x4 ){ 3674 sqlite3DebugPrintf( 3675 "Skip %s cost=%-3d,%3d order=%c", 3676 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, 3677 isOrdered>=0 ? isOrdered+'0' : '?'); 3678 sqlite3DebugPrintf(" vs %s cost=%-3d,%d order=%c\n", 3679 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 3680 pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 3681 } 3682 #endif 3683 /* Discard the candidate path from further consideration */ 3684 testcase( pTo->rCost==rCost ); 3685 continue; 3686 } 3687 testcase( pTo->rCost==rCost+1 ); 3688 /* Control reaches here if the candidate path is better than the 3689 ** pTo path. Replace pTo with the candidate. */ 3690 #ifdef WHERETRACE_ENABLED /* 0x4 */ 3691 if( sqlite3WhereTrace&0x4 ){ 3692 sqlite3DebugPrintf( 3693 "Update %s cost=%-3d,%3d order=%c", 3694 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, 3695 isOrdered>=0 ? isOrdered+'0' : '?'); 3696 sqlite3DebugPrintf(" was %s cost=%-3d,%3d order=%c\n", 3697 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 3698 pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 3699 } 3700 #endif 3701 } 3702 /* pWLoop is a winner. Add it to the set of best so far */ 3703 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 3704 pTo->revLoop = revMask; 3705 pTo->nRow = nOut; 3706 pTo->rCost = rCost; 3707 pTo->rUnsorted = rUnsorted; 3708 pTo->isOrdered = isOrdered; 3709 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 3710 pTo->aLoop[iLoop] = pWLoop; 3711 if( nTo>=mxChoice ){ 3712 mxI = 0; 3713 mxCost = aTo[0].rCost; 3714 mxUnsorted = aTo[0].nRow; 3715 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 3716 if( pTo->rCost>mxCost 3717 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 3718 ){ 3719 mxCost = pTo->rCost; 3720 mxUnsorted = pTo->rUnsorted; 3721 mxI = jj; 3722 } 3723 } 3724 } 3725 } 3726 } 3727 3728 #ifdef WHERETRACE_ENABLED /* >=2 */ 3729 if( sqlite3WhereTrace & 0x02 ){ 3730 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 3731 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 3732 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 3733 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 3734 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 3735 if( pTo->isOrdered>0 ){ 3736 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 3737 }else{ 3738 sqlite3DebugPrintf("\n"); 3739 } 3740 } 3741 } 3742 #endif 3743 3744 /* Swap the roles of aFrom and aTo for the next generation */ 3745 pFrom = aTo; 3746 aTo = aFrom; 3747 aFrom = pFrom; 3748 nFrom = nTo; 3749 } 3750 3751 if( nFrom==0 ){ 3752 sqlite3ErrorMsg(pParse, "no query solution"); 3753 sqlite3DbFree(db, pSpace); 3754 return SQLITE_ERROR; 3755 } 3756 3757 /* Find the lowest cost path. pFrom will be left pointing to that path */ 3758 pFrom = aFrom; 3759 for(ii=1; ii<nFrom; ii++){ 3760 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 3761 } 3762 assert( pWInfo->nLevel==nLoop ); 3763 /* Load the lowest cost path into pWInfo */ 3764 for(iLoop=0; iLoop<nLoop; iLoop++){ 3765 WhereLevel *pLevel = pWInfo->a + iLoop; 3766 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 3767 pLevel->iFrom = pWLoop->iTab; 3768 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 3769 } 3770 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 3771 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 3772 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 3773 && nRowEst 3774 ){ 3775 Bitmask notUsed; 3776 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 3777 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 3778 if( rc==pWInfo->pResultSet->nExpr ){ 3779 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 3780 } 3781 } 3782 if( pWInfo->pOrderBy ){ 3783 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 3784 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 3785 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 3786 } 3787 }else{ 3788 pWInfo->nOBSat = pFrom->isOrdered; 3789 if( pWInfo->nOBSat<0 ) pWInfo->nOBSat = 0; 3790 pWInfo->revMask = pFrom->revLoop; 3791 } 3792 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 3793 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 3794 ){ 3795 Bitmask revMask = 0; 3796 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 3797 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 3798 ); 3799 assert( pWInfo->sorted==0 ); 3800 if( nOrder==pWInfo->pOrderBy->nExpr ){ 3801 pWInfo->sorted = 1; 3802 pWInfo->revMask = revMask; 3803 } 3804 } 3805 } 3806 3807 3808 pWInfo->nRowOut = pFrom->nRow; 3809 3810 /* Free temporary memory and return success */ 3811 sqlite3DbFree(db, pSpace); 3812 return SQLITE_OK; 3813 } 3814 3815 /* 3816 ** Most queries use only a single table (they are not joins) and have 3817 ** simple == constraints against indexed fields. This routine attempts 3818 ** to plan those simple cases using much less ceremony than the 3819 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 3820 ** times for the common case. 3821 ** 3822 ** Return non-zero on success, if this query can be handled by this 3823 ** no-frills query planner. Return zero if this query needs the 3824 ** general-purpose query planner. 3825 */ 3826 static int whereShortCut(WhereLoopBuilder *pBuilder){ 3827 WhereInfo *pWInfo; 3828 struct SrcList_item *pItem; 3829 WhereClause *pWC; 3830 WhereTerm *pTerm; 3831 WhereLoop *pLoop; 3832 int iCur; 3833 int j; 3834 Table *pTab; 3835 Index *pIdx; 3836 3837 pWInfo = pBuilder->pWInfo; 3838 if( pWInfo->wctrlFlags & WHERE_FORCE_TABLE ) return 0; 3839 assert( pWInfo->pTabList->nSrc>=1 ); 3840 pItem = pWInfo->pTabList->a; 3841 pTab = pItem->pTab; 3842 if( IsVirtual(pTab) ) return 0; 3843 if( pItem->fg.isIndexedBy ) return 0; 3844 iCur = pItem->iCursor; 3845 pWC = &pWInfo->sWC; 3846 pLoop = pBuilder->pNew; 3847 pLoop->wsFlags = 0; 3848 pLoop->nSkip = 0; 3849 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0); 3850 if( pTerm ){ 3851 testcase( pTerm->eOperator & WO_IS ); 3852 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 3853 pLoop->aLTerm[0] = pTerm; 3854 pLoop->nLTerm = 1; 3855 pLoop->u.btree.nEq = 1; 3856 /* TUNING: Cost of a rowid lookup is 10 */ 3857 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 3858 }else{ 3859 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 3860 int opMask; 3861 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 3862 if( !IsUniqueIndex(pIdx) 3863 || pIdx->pPartIdxWhere!=0 3864 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 3865 ) continue; 3866 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 3867 for(j=0; j<pIdx->nKeyCol; j++){ 3868 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx); 3869 if( pTerm==0 ) break; 3870 testcase( pTerm->eOperator & WO_IS ); 3871 pLoop->aLTerm[j] = pTerm; 3872 } 3873 if( j!=pIdx->nKeyCol ) continue; 3874 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 3875 if( pIdx->isCovering || (pItem->colUsed & ~columnsInIndex(pIdx))==0 ){ 3876 pLoop->wsFlags |= WHERE_IDX_ONLY; 3877 } 3878 pLoop->nLTerm = j; 3879 pLoop->u.btree.nEq = j; 3880 pLoop->u.btree.pIndex = pIdx; 3881 /* TUNING: Cost of a unique index lookup is 15 */ 3882 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 3883 break; 3884 } 3885 } 3886 if( pLoop->wsFlags ){ 3887 pLoop->nOut = (LogEst)1; 3888 pWInfo->a[0].pWLoop = pLoop; 3889 pLoop->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); 3890 pWInfo->a[0].iTabCur = iCur; 3891 pWInfo->nRowOut = 1; 3892 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 3893 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 3894 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 3895 } 3896 #ifdef SQLITE_DEBUG 3897 pLoop->cId = '0'; 3898 #endif 3899 return 1; 3900 } 3901 return 0; 3902 } 3903 3904 /* 3905 ** Generate the beginning of the loop used for WHERE clause processing. 3906 ** The return value is a pointer to an opaque structure that contains 3907 ** information needed to terminate the loop. Later, the calling routine 3908 ** should invoke sqlite3WhereEnd() with the return value of this function 3909 ** in order to complete the WHERE clause processing. 3910 ** 3911 ** If an error occurs, this routine returns NULL. 3912 ** 3913 ** The basic idea is to do a nested loop, one loop for each table in 3914 ** the FROM clause of a select. (INSERT and UPDATE statements are the 3915 ** same as a SELECT with only a single table in the FROM clause.) For 3916 ** example, if the SQL is this: 3917 ** 3918 ** SELECT * FROM t1, t2, t3 WHERE ...; 3919 ** 3920 ** Then the code generated is conceptually like the following: 3921 ** 3922 ** foreach row1 in t1 do \ Code generated 3923 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 3924 ** foreach row3 in t3 do / 3925 ** ... 3926 ** end \ Code generated 3927 ** end |-- by sqlite3WhereEnd() 3928 ** end / 3929 ** 3930 ** Note that the loops might not be nested in the order in which they 3931 ** appear in the FROM clause if a different order is better able to make 3932 ** use of indices. Note also that when the IN operator appears in 3933 ** the WHERE clause, it might result in additional nested loops for 3934 ** scanning through all values on the right-hand side of the IN. 3935 ** 3936 ** There are Btree cursors associated with each table. t1 uses cursor 3937 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 3938 ** And so forth. This routine generates code to open those VDBE cursors 3939 ** and sqlite3WhereEnd() generates the code to close them. 3940 ** 3941 ** The code that sqlite3WhereBegin() generates leaves the cursors named 3942 ** in pTabList pointing at their appropriate entries. The [...] code 3943 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 3944 ** data from the various tables of the loop. 3945 ** 3946 ** If the WHERE clause is empty, the foreach loops must each scan their 3947 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 3948 ** the tables have indices and there are terms in the WHERE clause that 3949 ** refer to those indices, a complete table scan can be avoided and the 3950 ** code will run much faster. Most of the work of this routine is checking 3951 ** to see if there are indices that can be used to speed up the loop. 3952 ** 3953 ** Terms of the WHERE clause are also used to limit which rows actually 3954 ** make it to the "..." in the middle of the loop. After each "foreach", 3955 ** terms of the WHERE clause that use only terms in that loop and outer 3956 ** loops are evaluated and if false a jump is made around all subsequent 3957 ** inner loops (or around the "..." if the test occurs within the inner- 3958 ** most loop) 3959 ** 3960 ** OUTER JOINS 3961 ** 3962 ** An outer join of tables t1 and t2 is conceptally coded as follows: 3963 ** 3964 ** foreach row1 in t1 do 3965 ** flag = 0 3966 ** foreach row2 in t2 do 3967 ** start: 3968 ** ... 3969 ** flag = 1 3970 ** end 3971 ** if flag==0 then 3972 ** move the row2 cursor to a null row 3973 ** goto start 3974 ** fi 3975 ** end 3976 ** 3977 ** ORDER BY CLAUSE PROCESSING 3978 ** 3979 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 3980 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 3981 ** if there is one. If there is no ORDER BY clause or if this routine 3982 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 3983 ** 3984 ** The iIdxCur parameter is the cursor number of an index. If 3985 ** WHERE_ONETABLE_ONLY is set, iIdxCur is the cursor number of an index 3986 ** to use for OR clause processing. The WHERE clause should use this 3987 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 3988 ** the first cursor in an array of cursors for all indices. iIdxCur should 3989 ** be used to compute the appropriate cursor depending on which index is 3990 ** used. 3991 */ 3992 WhereInfo *sqlite3WhereBegin( 3993 Parse *pParse, /* The parser context */ 3994 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 3995 Expr *pWhere, /* The WHERE clause */ 3996 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 3997 ExprList *pResultSet, /* Result set of the query */ 3998 u16 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */ 3999 int iIdxCur /* If WHERE_ONETABLE_ONLY is set, index cursor number */ 4000 ){ 4001 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4002 int nTabList; /* Number of elements in pTabList */ 4003 WhereInfo *pWInfo; /* Will become the return value of this function */ 4004 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4005 Bitmask notReady; /* Cursors that are not yet positioned */ 4006 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 4007 WhereMaskSet *pMaskSet; /* The expression mask set */ 4008 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 4009 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 4010 int ii; /* Loop counter */ 4011 sqlite3 *db; /* Database connection */ 4012 int rc; /* Return code */ 4013 u8 bFordelete = 0; 4014 4015 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 4016 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 4017 && (wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 4018 )); 4019 4020 /* Variable initialization */ 4021 db = pParse->db; 4022 memset(&sWLB, 0, sizeof(sWLB)); 4023 4024 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 4025 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 4026 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 4027 sWLB.pOrderBy = pOrderBy; 4028 4029 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 4030 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 4031 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 4032 wctrlFlags &= ~WHERE_WANT_DISTINCT; 4033 } 4034 4035 /* The number of tables in the FROM clause is limited by the number of 4036 ** bits in a Bitmask 4037 */ 4038 testcase( pTabList->nSrc==BMS ); 4039 if( pTabList->nSrc>BMS ){ 4040 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 4041 return 0; 4042 } 4043 4044 /* This function normally generates a nested loop for all tables in 4045 ** pTabList. But if the WHERE_ONETABLE_ONLY flag is set, then we should 4046 ** only generate code for the first table in pTabList and assume that 4047 ** any cursors associated with subsequent tables are uninitialized. 4048 */ 4049 nTabList = (wctrlFlags & WHERE_ONETABLE_ONLY) ? 1 : pTabList->nSrc; 4050 4051 /* Allocate and initialize the WhereInfo structure that will become the 4052 ** return value. A single allocation is used to store the WhereInfo 4053 ** struct, the contents of WhereInfo.a[], the WhereClause structure 4054 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 4055 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 4056 ** some architectures. Hence the ROUND8() below. 4057 */ 4058 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 4059 pWInfo = sqlite3DbMallocZero(db, nByteWInfo + sizeof(WhereLoop)); 4060 if( db->mallocFailed ){ 4061 sqlite3DbFree(db, pWInfo); 4062 pWInfo = 0; 4063 goto whereBeginError; 4064 } 4065 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 4066 pWInfo->nLevel = nTabList; 4067 pWInfo->pParse = pParse; 4068 pWInfo->pTabList = pTabList; 4069 pWInfo->pOrderBy = pOrderBy; 4070 pWInfo->pResultSet = pResultSet; 4071 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(v); 4072 pWInfo->wctrlFlags = wctrlFlags; 4073 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 4074 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 4075 pMaskSet = &pWInfo->sMaskSet; 4076 sWLB.pWInfo = pWInfo; 4077 sWLB.pWC = &pWInfo->sWC; 4078 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 4079 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 4080 whereLoopInit(sWLB.pNew); 4081 #ifdef SQLITE_DEBUG 4082 sWLB.pNew->cId = '*'; 4083 #endif 4084 4085 /* Split the WHERE clause into separate subexpressions where each 4086 ** subexpression is separated by an AND operator. 4087 */ 4088 initMaskSet(pMaskSet); 4089 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 4090 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 4091 4092 /* Special case: a WHERE clause that is constant. Evaluate the 4093 ** expression and either jump over all of the code or fall thru. 4094 */ 4095 for(ii=0; ii<sWLB.pWC->nTerm; ii++){ 4096 if( nTabList==0 || sqlite3ExprIsConstantNotJoin(sWLB.pWC->a[ii].pExpr) ){ 4097 sqlite3ExprIfFalse(pParse, sWLB.pWC->a[ii].pExpr, pWInfo->iBreak, 4098 SQLITE_JUMPIFNULL); 4099 sWLB.pWC->a[ii].wtFlags |= TERM_CODED; 4100 } 4101 } 4102 4103 /* Special case: No FROM clause 4104 */ 4105 if( nTabList==0 ){ 4106 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 4107 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4108 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4109 } 4110 } 4111 4112 /* Assign a bit from the bitmask to every term in the FROM clause. 4113 ** 4114 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 4115 ** 4116 ** The rule of the previous sentence ensures thta if X is the bitmask for 4117 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 4118 ** Knowing the bitmask for all tables to the left of a left join is 4119 ** important. Ticket #3015. 4120 ** 4121 ** Note that bitmasks are created for all pTabList->nSrc tables in 4122 ** pTabList, not just the first nTabList tables. nTabList is normally 4123 ** equal to pTabList->nSrc but might be shortened to 1 if the 4124 ** WHERE_ONETABLE_ONLY flag is set. 4125 */ 4126 for(ii=0; ii<pTabList->nSrc; ii++){ 4127 createMask(pMaskSet, pTabList->a[ii].iCursor); 4128 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 4129 } 4130 #ifdef SQLITE_DEBUG 4131 for(ii=0; ii<pTabList->nSrc; ii++){ 4132 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 4133 assert( m==MASKBIT(ii) ); 4134 } 4135 #endif 4136 4137 /* Analyze all of the subexpressions. */ 4138 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 4139 if( db->mallocFailed ) goto whereBeginError; 4140 4141 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4142 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 4143 /* The DISTINCT marking is pointless. Ignore it. */ 4144 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4145 }else if( pOrderBy==0 ){ 4146 /* Try to ORDER BY the result set to make distinct processing easier */ 4147 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 4148 pWInfo->pOrderBy = pResultSet; 4149 } 4150 } 4151 4152 /* Construct the WhereLoop objects */ 4153 WHERETRACE(0xffff,("*** Optimizer Start *** (wctrlFlags: 0x%x)\n", 4154 wctrlFlags)); 4155 #if defined(WHERETRACE_ENABLED) 4156 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 4157 int i; 4158 for(i=0; i<sWLB.pWC->nTerm; i++){ 4159 whereTermPrint(&sWLB.pWC->a[i], i); 4160 } 4161 } 4162 #endif 4163 4164 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 4165 rc = whereLoopAddAll(&sWLB); 4166 if( rc ) goto whereBeginError; 4167 4168 #ifdef WHERETRACE_ENABLED 4169 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4170 WhereLoop *p; 4171 int i; 4172 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4173 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4174 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4175 p->cId = zLabel[i%sizeof(zLabel)]; 4176 whereLoopPrint(p, sWLB.pWC); 4177 } 4178 } 4179 #endif 4180 4181 wherePathSolver(pWInfo, 0); 4182 if( db->mallocFailed ) goto whereBeginError; 4183 if( pWInfo->pOrderBy ){ 4184 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 4185 if( db->mallocFailed ) goto whereBeginError; 4186 } 4187 } 4188 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 4189 pWInfo->revMask = (Bitmask)(-1); 4190 } 4191 if( pParse->nErr || NEVER(db->mallocFailed) ){ 4192 goto whereBeginError; 4193 } 4194 #ifdef WHERETRACE_ENABLED 4195 if( sqlite3WhereTrace ){ 4196 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 4197 if( pWInfo->nOBSat>0 ){ 4198 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 4199 } 4200 switch( pWInfo->eDistinct ){ 4201 case WHERE_DISTINCT_UNIQUE: { 4202 sqlite3DebugPrintf(" DISTINCT=unique"); 4203 break; 4204 } 4205 case WHERE_DISTINCT_ORDERED: { 4206 sqlite3DebugPrintf(" DISTINCT=ordered"); 4207 break; 4208 } 4209 case WHERE_DISTINCT_UNORDERED: { 4210 sqlite3DebugPrintf(" DISTINCT=unordered"); 4211 break; 4212 } 4213 } 4214 sqlite3DebugPrintf("\n"); 4215 for(ii=0; ii<pWInfo->nLevel; ii++){ 4216 whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 4217 } 4218 } 4219 #endif 4220 /* Attempt to omit tables from the join that do not effect the result */ 4221 if( pWInfo->nLevel>=2 4222 && pResultSet!=0 4223 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 4224 ){ 4225 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet); 4226 if( sWLB.pOrderBy ){ 4227 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy); 4228 } 4229 while( pWInfo->nLevel>=2 ){ 4230 WhereTerm *pTerm, *pEnd; 4231 pLoop = pWInfo->a[pWInfo->nLevel-1].pWLoop; 4232 if( (pWInfo->pTabList->a[pLoop->iTab].fg.jointype & JT_LEFT)==0 ) break; 4233 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 4234 && (pLoop->wsFlags & WHERE_ONEROW)==0 4235 ){ 4236 break; 4237 } 4238 if( (tabUsed & pLoop->maskSelf)!=0 ) break; 4239 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; 4240 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4241 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 4242 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 4243 ){ 4244 break; 4245 } 4246 } 4247 if( pTerm<pEnd ) break; 4248 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 4249 pWInfo->nLevel--; 4250 nTabList--; 4251 } 4252 } 4253 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 4254 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 4255 4256 /* If the caller is an UPDATE or DELETE statement that is requesting 4257 ** to use a one-pass algorithm, determine if this is appropriate. 4258 ** The one-pass algorithm only works if the WHERE clause constrains 4259 ** the statement to update or delete a single row. 4260 */ 4261 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 4262 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 4263 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 4264 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 4265 if( bOnerow || ( (wctrlFlags & WHERE_ONEPASS_MULTIROW) 4266 && 0==(wsFlags & WHERE_VIRTUALTABLE) 4267 )){ 4268 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 4269 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 4270 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 4271 bFordelete = OPFLAG_FORDELETE; 4272 } 4273 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 4274 } 4275 } 4276 } 4277 4278 /* Open all tables in the pTabList and any indices selected for 4279 ** searching those tables. 4280 */ 4281 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 4282 Table *pTab; /* Table to open */ 4283 int iDb; /* Index of database containing table/index */ 4284 struct SrcList_item *pTabItem; 4285 4286 pTabItem = &pTabList->a[pLevel->iFrom]; 4287 pTab = pTabItem->pTab; 4288 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 4289 pLoop = pLevel->pWLoop; 4290 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 4291 /* Do nothing */ 4292 }else 4293 #ifndef SQLITE_OMIT_VIRTUALTABLE 4294 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 4295 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 4296 int iCur = pTabItem->iCursor; 4297 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 4298 }else if( IsVirtual(pTab) ){ 4299 /* noop */ 4300 }else 4301 #endif 4302 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 4303 && (wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 ){ 4304 int op = OP_OpenRead; 4305 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4306 op = OP_OpenWrite; 4307 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 4308 }; 4309 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 4310 assert( pTabItem->iCursor==pLevel->iTabCur ); 4311 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 4312 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 4313 if( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol<BMS && HasRowid(pTab) ){ 4314 Bitmask b = pTabItem->colUsed; 4315 int n = 0; 4316 for(; b; b=b>>1, n++){} 4317 sqlite3VdbeChangeP4(v, sqlite3VdbeCurrentAddr(v)-1, 4318 SQLITE_INT_TO_PTR(n), P4_INT32); 4319 assert( n<=pTab->nCol ); 4320 } 4321 #ifdef SQLITE_ENABLE_CURSOR_HINTS 4322 if( pLoop->u.btree.pIndex!=0 ){ 4323 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 4324 }else 4325 #endif 4326 { 4327 sqlite3VdbeChangeP5(v, bFordelete); 4328 } 4329 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 4330 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 4331 (const u8*)&pTabItem->colUsed, P4_INT64); 4332 #endif 4333 }else{ 4334 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 4335 } 4336 if( pLoop->wsFlags & WHERE_INDEXED ){ 4337 Index *pIx = pLoop->u.btree.pIndex; 4338 int iIndexCur; 4339 int op = OP_OpenRead; 4340 /* iIdxCur is always set if to a positive value if ONEPASS is possible */ 4341 assert( iIdxCur!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 4342 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 4343 && (wctrlFlags & WHERE_ONETABLE_ONLY)!=0 4344 ){ 4345 /* This is one term of an OR-optimization using the PRIMARY KEY of a 4346 ** WITHOUT ROWID table. No need for a separate index */ 4347 iIndexCur = pLevel->iTabCur; 4348 op = 0; 4349 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4350 Index *pJ = pTabItem->pTab->pIndex; 4351 iIndexCur = iIdxCur; 4352 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 4353 while( ALWAYS(pJ) && pJ!=pIx ){ 4354 iIndexCur++; 4355 pJ = pJ->pNext; 4356 } 4357 op = OP_OpenWrite; 4358 pWInfo->aiCurOnePass[1] = iIndexCur; 4359 }else if( iIdxCur && (wctrlFlags & WHERE_ONETABLE_ONLY)!=0 ){ 4360 iIndexCur = iIdxCur; 4361 if( wctrlFlags & WHERE_REOPEN_IDX ) op = OP_ReopenIdx; 4362 }else{ 4363 iIndexCur = pParse->nTab++; 4364 } 4365 pLevel->iIdxCur = iIndexCur; 4366 assert( pIx->pSchema==pTab->pSchema ); 4367 assert( iIndexCur>=0 ); 4368 if( op ){ 4369 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 4370 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 4371 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 4372 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 4373 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 4374 ){ 4375 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); /* Hint to COMDB2 */ 4376 } 4377 VdbeComment((v, "%s", pIx->zName)); 4378 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 4379 { 4380 u64 colUsed = 0; 4381 int ii, jj; 4382 for(ii=0; ii<pIx->nColumn; ii++){ 4383 jj = pIx->aiColumn[ii]; 4384 if( jj<0 ) continue; 4385 if( jj>63 ) jj = 63; 4386 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 4387 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 4388 } 4389 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 4390 (u8*)&colUsed, P4_INT64); 4391 } 4392 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 4393 } 4394 } 4395 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 4396 } 4397 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 4398 if( db->mallocFailed ) goto whereBeginError; 4399 4400 /* Generate the code to do the search. Each iteration of the for 4401 ** loop below generates code for a single nested loop of the VM 4402 ** program. 4403 */ 4404 notReady = ~(Bitmask)0; 4405 for(ii=0; ii<nTabList; ii++){ 4406 int addrExplain; 4407 int wsFlags; 4408 pLevel = &pWInfo->a[ii]; 4409 wsFlags = pLevel->pWLoop->wsFlags; 4410 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 4411 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 4412 constructAutomaticIndex(pParse, &pWInfo->sWC, 4413 &pTabList->a[pLevel->iFrom], notReady, pLevel); 4414 if( db->mallocFailed ) goto whereBeginError; 4415 } 4416 #endif 4417 addrExplain = sqlite3WhereExplainOneScan( 4418 pParse, pTabList, pLevel, ii, pLevel->iFrom, wctrlFlags 4419 ); 4420 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 4421 notReady = sqlite3WhereCodeOneLoopStart(pWInfo, ii, notReady); 4422 pWInfo->iContinue = pLevel->addrCont; 4423 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_ONETABLE_ONLY)==0 ){ 4424 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 4425 } 4426 } 4427 4428 /* Done. */ 4429 VdbeModuleComment((v, "Begin WHERE-core")); 4430 return pWInfo; 4431 4432 /* Jump here if malloc fails */ 4433 whereBeginError: 4434 if( pWInfo ){ 4435 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 4436 whereInfoFree(db, pWInfo); 4437 } 4438 return 0; 4439 } 4440 4441 /* 4442 ** Generate the end of the WHERE loop. See comments on 4443 ** sqlite3WhereBegin() for additional information. 4444 */ 4445 void sqlite3WhereEnd(WhereInfo *pWInfo){ 4446 Parse *pParse = pWInfo->pParse; 4447 Vdbe *v = pParse->pVdbe; 4448 int i; 4449 WhereLevel *pLevel; 4450 WhereLoop *pLoop; 4451 SrcList *pTabList = pWInfo->pTabList; 4452 sqlite3 *db = pParse->db; 4453 4454 /* Generate loop termination code. 4455 */ 4456 VdbeModuleComment((v, "End WHERE-core")); 4457 sqlite3ExprCacheClear(pParse); 4458 for(i=pWInfo->nLevel-1; i>=0; i--){ 4459 int addr; 4460 pLevel = &pWInfo->a[i]; 4461 pLoop = pLevel->pWLoop; 4462 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 4463 if( pLevel->op!=OP_Noop ){ 4464 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 4465 sqlite3VdbeChangeP5(v, pLevel->p5); 4466 VdbeCoverage(v); 4467 VdbeCoverageIf(v, pLevel->op==OP_Next); 4468 VdbeCoverageIf(v, pLevel->op==OP_Prev); 4469 VdbeCoverageIf(v, pLevel->op==OP_VNext); 4470 } 4471 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 4472 struct InLoop *pIn; 4473 int j; 4474 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 4475 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 4476 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 4477 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 4478 VdbeCoverage(v); 4479 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_PrevIfOpen); 4480 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_NextIfOpen); 4481 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 4482 } 4483 } 4484 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 4485 if( pLevel->addrSkip ){ 4486 sqlite3VdbeGoto(v, pLevel->addrSkip); 4487 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 4488 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 4489 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 4490 } 4491 if( pLevel->addrLikeRep ){ 4492 int op; 4493 if( sqlite3VdbeGetOp(v, pLevel->addrLikeRep-1)->p1 ){ 4494 op = OP_DecrJumpZero; 4495 }else{ 4496 op = OP_JumpZeroIncr; 4497 } 4498 sqlite3VdbeAddOp2(v, op, pLevel->iLikeRepCntr, pLevel->addrLikeRep); 4499 VdbeCoverage(v); 4500 } 4501 if( pLevel->iLeftJoin ){ 4502 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 4503 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 4504 || (pLoop->wsFlags & WHERE_INDEXED)!=0 ); 4505 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 ){ 4506 sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor); 4507 } 4508 if( pLoop->wsFlags & WHERE_INDEXED ){ 4509 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 4510 } 4511 if( pLevel->op==OP_Return ){ 4512 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 4513 }else{ 4514 sqlite3VdbeGoto(v, pLevel->addrFirst); 4515 } 4516 sqlite3VdbeJumpHere(v, addr); 4517 } 4518 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 4519 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 4520 } 4521 4522 /* The "break" point is here, just past the end of the outer loop. 4523 ** Set it. 4524 */ 4525 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 4526 4527 assert( pWInfo->nLevel<=pTabList->nSrc ); 4528 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 4529 int k, last; 4530 VdbeOp *pOp; 4531 Index *pIdx = 0; 4532 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; 4533 Table *pTab = pTabItem->pTab; 4534 assert( pTab!=0 ); 4535 pLoop = pLevel->pWLoop; 4536 4537 /* For a co-routine, change all OP_Column references to the table of 4538 ** the co-routine into OP_Copy of result contained in a register. 4539 ** OP_Rowid becomes OP_Null. 4540 */ 4541 if( pTabItem->fg.viaCoroutine && !db->mallocFailed ){ 4542 translateColumnToCopy(v, pLevel->addrBody, pLevel->iTabCur, 4543 pTabItem->regResult, 0); 4544 continue; 4545 } 4546 4547 /* Close all of the cursors that were opened by sqlite3WhereBegin. 4548 ** Except, do not close cursors that will be reused by the OR optimization 4549 ** (WHERE_OMIT_OPEN_CLOSE). And do not close the OP_OpenWrite cursors 4550 ** created for the ONEPASS optimization. 4551 */ 4552 if( (pTab->tabFlags & TF_Ephemeral)==0 4553 && pTab->pSelect==0 4554 && (pWInfo->wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 4555 ){ 4556 int ws = pLoop->wsFlags; 4557 if( pWInfo->eOnePass==ONEPASS_OFF && (ws & WHERE_IDX_ONLY)==0 ){ 4558 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); 4559 } 4560 if( (ws & WHERE_INDEXED)!=0 4561 && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0 4562 && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1] 4563 ){ 4564 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); 4565 } 4566 } 4567 4568 /* If this scan uses an index, make VDBE code substitutions to read data 4569 ** from the index instead of from the table where possible. In some cases 4570 ** this optimization prevents the table from ever being read, which can 4571 ** yield a significant performance boost. 4572 ** 4573 ** Calls to the code generator in between sqlite3WhereBegin and 4574 ** sqlite3WhereEnd will have created code that references the table 4575 ** directly. This loop scans all that code looking for opcodes 4576 ** that reference the table and converts them into opcodes that 4577 ** reference the index. 4578 */ 4579 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 4580 pIdx = pLoop->u.btree.pIndex; 4581 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 4582 pIdx = pLevel->u.pCovidx; 4583 } 4584 if( pIdx 4585 && (pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable)) 4586 && !db->mallocFailed 4587 ){ 4588 last = sqlite3VdbeCurrentAddr(v); 4589 k = pLevel->addrBody; 4590 pOp = sqlite3VdbeGetOp(v, k); 4591 for(; k<last; k++, pOp++){ 4592 if( pOp->p1!=pLevel->iTabCur ) continue; 4593 if( pOp->opcode==OP_Column ){ 4594 int x = pOp->p2; 4595 assert( pIdx->pTable==pTab ); 4596 if( !HasRowid(pTab) ){ 4597 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 4598 x = pPk->aiColumn[x]; 4599 assert( x>=0 ); 4600 } 4601 x = sqlite3ColumnOfIndex(pIdx, x); 4602 if( x>=0 ){ 4603 pOp->p2 = x; 4604 pOp->p1 = pLevel->iIdxCur; 4605 } 4606 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 ); 4607 }else if( pOp->opcode==OP_Rowid ){ 4608 pOp->p1 = pLevel->iIdxCur; 4609 pOp->opcode = OP_IdxRowid; 4610 } 4611 } 4612 } 4613 } 4614 4615 /* Final cleanup 4616 */ 4617 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 4618 whereInfoFree(db, pWInfo); 4619 return; 4620 } 4621