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