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 LogEst sqlite3WhereOutputRowCount(WhereInfo *pWInfo){ 35 return 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 TRUE if the innermost loop of the WHERE clause implementation 56 ** returns rows in ORDER BY order for complete run of the inner loop. 57 ** 58 ** Across multiple iterations of outer loops, the output rows need not be 59 ** sorted. As long as rows are sorted for just the innermost loop, this 60 ** routine can return TRUE. 61 */ 62 int sqlite3WhereOrderedInnerLoop(WhereInfo *pWInfo){ 63 return pWInfo->bOrderedInnerLoop; 64 } 65 66 /* 67 ** Return the VDBE address or label to jump to in order to continue 68 ** immediately with the next row of a WHERE clause. 69 */ 70 int sqlite3WhereContinueLabel(WhereInfo *pWInfo){ 71 assert( pWInfo->iContinue!=0 ); 72 return pWInfo->iContinue; 73 } 74 75 /* 76 ** Return the VDBE address or label to jump to in order to break 77 ** out of a WHERE loop. 78 */ 79 int sqlite3WhereBreakLabel(WhereInfo *pWInfo){ 80 return pWInfo->iBreak; 81 } 82 83 /* 84 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to 85 ** operate directly on the rowis returned by a WHERE clause. Return 86 ** ONEPASS_SINGLE (1) if the statement can operation directly because only 87 ** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass 88 ** optimization can be used on multiple 89 ** 90 ** If the ONEPASS optimization is used (if this routine returns true) 91 ** then also write the indices of open cursors used by ONEPASS 92 ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data 93 ** table and iaCur[1] gets the cursor used by an auxiliary index. 94 ** Either value may be -1, indicating that cursor is not used. 95 ** Any cursors returned will have been opened for writing. 96 ** 97 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is 98 ** unable to use the ONEPASS optimization. 99 */ 100 int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){ 101 memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2); 102 #ifdef WHERETRACE_ENABLED 103 if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){ 104 sqlite3DebugPrintf("%s cursors: %d %d\n", 105 pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI", 106 aiCur[0], aiCur[1]); 107 } 108 #endif 109 return pWInfo->eOnePass; 110 } 111 112 /* 113 ** Move the content of pSrc into pDest 114 */ 115 static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){ 116 pDest->n = pSrc->n; 117 memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0])); 118 } 119 120 /* 121 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet. 122 ** 123 ** The new entry might overwrite an existing entry, or it might be 124 ** appended, or it might be discarded. Do whatever is the right thing 125 ** so that pSet keeps the N_OR_COST best entries seen so far. 126 */ 127 static int whereOrInsert( 128 WhereOrSet *pSet, /* The WhereOrSet to be updated */ 129 Bitmask prereq, /* Prerequisites of the new entry */ 130 LogEst rRun, /* Run-cost of the new entry */ 131 LogEst nOut /* Number of outputs for the new entry */ 132 ){ 133 u16 i; 134 WhereOrCost *p; 135 for(i=pSet->n, p=pSet->a; i>0; i--, p++){ 136 if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){ 137 goto whereOrInsert_done; 138 } 139 if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){ 140 return 0; 141 } 142 } 143 if( pSet->n<N_OR_COST ){ 144 p = &pSet->a[pSet->n++]; 145 p->nOut = nOut; 146 }else{ 147 p = pSet->a; 148 for(i=1; i<pSet->n; i++){ 149 if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i; 150 } 151 if( p->rRun<=rRun ) return 0; 152 } 153 whereOrInsert_done: 154 p->prereq = prereq; 155 p->rRun = rRun; 156 if( p->nOut>nOut ) p->nOut = nOut; 157 return 1; 158 } 159 160 /* 161 ** Return the bitmask for the given cursor number. Return 0 if 162 ** iCursor is not in the set. 163 */ 164 Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){ 165 int i; 166 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 ); 167 for(i=0; i<pMaskSet->n; i++){ 168 if( pMaskSet->ix[i]==iCursor ){ 169 return MASKBIT(i); 170 } 171 } 172 return 0; 173 } 174 175 /* 176 ** Create a new mask for cursor iCursor. 177 ** 178 ** There is one cursor per table in the FROM clause. The number of 179 ** tables in the FROM clause is limited by a test early in the 180 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] 181 ** array will never overflow. 182 */ 183 static void createMask(WhereMaskSet *pMaskSet, int iCursor){ 184 assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); 185 pMaskSet->ix[pMaskSet->n++] = iCursor; 186 } 187 188 /* 189 ** Advance to the next WhereTerm that matches according to the criteria 190 ** established when the pScan object was initialized by whereScanInit(). 191 ** Return NULL if there are no more matching WhereTerms. 192 */ 193 static WhereTerm *whereScanNext(WhereScan *pScan){ 194 int iCur; /* The cursor on the LHS of the term */ 195 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */ 196 Expr *pX; /* An expression being tested */ 197 WhereClause *pWC; /* Shorthand for pScan->pWC */ 198 WhereTerm *pTerm; /* The term being tested */ 199 int k = pScan->k; /* Where to start scanning */ 200 201 assert( pScan->iEquiv<=pScan->nEquiv ); 202 pWC = pScan->pWC; 203 while(1){ 204 iColumn = pScan->aiColumn[pScan->iEquiv-1]; 205 iCur = pScan->aiCur[pScan->iEquiv-1]; 206 assert( pWC!=0 ); 207 do{ 208 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){ 209 if( pTerm->leftCursor==iCur 210 && pTerm->u.leftColumn==iColumn 211 && (iColumn!=XN_EXPR 212 || sqlite3ExprCompareSkip(pTerm->pExpr->pLeft, 213 pScan->pIdxExpr,iCur)==0) 214 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin)) 215 ){ 216 if( (pTerm->eOperator & WO_EQUIV)!=0 217 && pScan->nEquiv<ArraySize(pScan->aiCur) 218 && (pX = sqlite3ExprSkipCollate(pTerm->pExpr->pRight))->op==TK_COLUMN 219 ){ 220 int j; 221 for(j=0; j<pScan->nEquiv; j++){ 222 if( pScan->aiCur[j]==pX->iTable 223 && pScan->aiColumn[j]==pX->iColumn ){ 224 break; 225 } 226 } 227 if( j==pScan->nEquiv ){ 228 pScan->aiCur[j] = pX->iTable; 229 pScan->aiColumn[j] = pX->iColumn; 230 pScan->nEquiv++; 231 } 232 } 233 if( (pTerm->eOperator & pScan->opMask)!=0 ){ 234 /* Verify the affinity and collating sequence match */ 235 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){ 236 CollSeq *pColl; 237 Parse *pParse = pWC->pWInfo->pParse; 238 pX = pTerm->pExpr; 239 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){ 240 continue; 241 } 242 assert(pX->pLeft); 243 pColl = sqlite3BinaryCompareCollSeq(pParse, 244 pX->pLeft, pX->pRight); 245 if( pColl==0 ) pColl = pParse->db->pDfltColl; 246 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){ 247 continue; 248 } 249 } 250 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0 251 && (pX = pTerm->pExpr->pRight)->op==TK_COLUMN 252 && pX->iTable==pScan->aiCur[0] 253 && pX->iColumn==pScan->aiColumn[0] 254 ){ 255 testcase( pTerm->eOperator & WO_IS ); 256 continue; 257 } 258 pScan->pWC = pWC; 259 pScan->k = k+1; 260 return pTerm; 261 } 262 } 263 } 264 pWC = pWC->pOuter; 265 k = 0; 266 }while( pWC!=0 ); 267 if( pScan->iEquiv>=pScan->nEquiv ) break; 268 pWC = pScan->pOrigWC; 269 k = 0; 270 pScan->iEquiv++; 271 } 272 return 0; 273 } 274 275 /* 276 ** Initialize a WHERE clause scanner object. Return a pointer to the 277 ** first match. Return NULL if there are no matches. 278 ** 279 ** The scanner will be searching the WHERE clause pWC. It will look 280 ** for terms of the form "X <op> <expr>" where X is column iColumn of table 281 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx 282 ** must be one of the indexes of table iCur. 283 ** 284 ** The <op> must be one of the operators described by opMask. 285 ** 286 ** If the search is for X and the WHERE clause contains terms of the 287 ** form X=Y then this routine might also return terms of the form 288 ** "Y <op> <expr>". The number of levels of transitivity is limited, 289 ** but is enough to handle most commonly occurring SQL statements. 290 ** 291 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with 292 ** index pIdx. 293 */ 294 static WhereTerm *whereScanInit( 295 WhereScan *pScan, /* The WhereScan object being initialized */ 296 WhereClause *pWC, /* The WHERE clause to be scanned */ 297 int iCur, /* Cursor to scan for */ 298 int iColumn, /* Column to scan for */ 299 u32 opMask, /* Operator(s) to scan for */ 300 Index *pIdx /* Must be compatible with this index */ 301 ){ 302 pScan->pOrigWC = pWC; 303 pScan->pWC = pWC; 304 pScan->pIdxExpr = 0; 305 pScan->idxaff = 0; 306 pScan->zCollName = 0; 307 if( pIdx ){ 308 int j = iColumn; 309 iColumn = pIdx->aiColumn[j]; 310 if( iColumn==XN_EXPR ){ 311 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr; 312 pScan->zCollName = pIdx->azColl[j]; 313 }else if( iColumn==pIdx->pTable->iPKey ){ 314 iColumn = XN_ROWID; 315 }else if( iColumn>=0 ){ 316 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; 317 pScan->zCollName = pIdx->azColl[j]; 318 } 319 }else if( iColumn==XN_EXPR ){ 320 return 0; 321 } 322 pScan->opMask = opMask; 323 pScan->k = 0; 324 pScan->aiCur[0] = iCur; 325 pScan->aiColumn[0] = iColumn; 326 pScan->nEquiv = 1; 327 pScan->iEquiv = 1; 328 return whereScanNext(pScan); 329 } 330 331 /* 332 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 333 ** where X is a reference to the iColumn of table iCur or of index pIdx 334 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by 335 ** the op parameter. Return a pointer to the term. Return 0 if not found. 336 ** 337 ** If pIdx!=0 then it must be one of the indexes of table iCur. 338 ** Search for terms matching the iColumn-th column of pIdx 339 ** rather than the iColumn-th column of table iCur. 340 ** 341 ** The term returned might by Y=<expr> if there is another constraint in 342 ** the WHERE clause that specifies that X=Y. Any such constraints will be 343 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 344 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11 345 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10 346 ** other equivalent values. Hence a search for X will return <expr> if X=A1 347 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>. 348 ** 349 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 350 ** then try for the one with no dependencies on <expr> - in other words where 351 ** <expr> is a constant expression of some kind. Only return entries of 352 ** the form "X <op> Y" where Y is a column in another table if no terms of 353 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 354 ** exist, try to return a term that does not use WO_EQUIV. 355 */ 356 WhereTerm *sqlite3WhereFindTerm( 357 WhereClause *pWC, /* The WHERE clause to be searched */ 358 int iCur, /* Cursor number of LHS */ 359 int iColumn, /* Column number of LHS */ 360 Bitmask notReady, /* RHS must not overlap with this mask */ 361 u32 op, /* Mask of WO_xx values describing operator */ 362 Index *pIdx /* Must be compatible with this index, if not NULL */ 363 ){ 364 WhereTerm *pResult = 0; 365 WhereTerm *p; 366 WhereScan scan; 367 368 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); 369 op &= WO_EQ|WO_IS; 370 while( p ){ 371 if( (p->prereqRight & notReady)==0 ){ 372 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){ 373 testcase( p->eOperator & WO_IS ); 374 return p; 375 } 376 if( pResult==0 ) pResult = p; 377 } 378 p = whereScanNext(&scan); 379 } 380 return pResult; 381 } 382 383 /* 384 ** This function searches pList for an entry that matches the iCol-th column 385 ** of index pIdx. 386 ** 387 ** If such an expression is found, its index in pList->a[] is returned. If 388 ** no expression is found, -1 is returned. 389 */ 390 static int findIndexCol( 391 Parse *pParse, /* Parse context */ 392 ExprList *pList, /* Expression list to search */ 393 int iBase, /* Cursor for table associated with pIdx */ 394 Index *pIdx, /* Index to match column of */ 395 int iCol /* Column of index to match */ 396 ){ 397 int i; 398 const char *zColl = pIdx->azColl[iCol]; 399 400 for(i=0; i<pList->nExpr; i++){ 401 Expr *p = sqlite3ExprSkipCollate(pList->a[i].pExpr); 402 if( p->op==TK_COLUMN 403 && p->iColumn==pIdx->aiColumn[iCol] 404 && p->iTable==iBase 405 ){ 406 CollSeq *pColl = sqlite3ExprCollSeq(pParse, pList->a[i].pExpr); 407 if( pColl && 0==sqlite3StrICmp(pColl->zName, zColl) ){ 408 return i; 409 } 410 } 411 } 412 413 return -1; 414 } 415 416 /* 417 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL 418 */ 419 static int indexColumnNotNull(Index *pIdx, int iCol){ 420 int j; 421 assert( pIdx!=0 ); 422 assert( iCol>=0 && iCol<pIdx->nColumn ); 423 j = pIdx->aiColumn[iCol]; 424 if( j>=0 ){ 425 return pIdx->pTable->aCol[j].notNull; 426 }else if( j==(-1) ){ 427 return 1; 428 }else{ 429 assert( j==(-2) ); 430 return 0; /* Assume an indexed expression can always yield a NULL */ 431 432 } 433 } 434 435 /* 436 ** Return true if the DISTINCT expression-list passed as the third argument 437 ** is redundant. 438 ** 439 ** A DISTINCT list is redundant if any subset of the columns in the 440 ** DISTINCT list are collectively unique and individually non-null. 441 */ 442 static int isDistinctRedundant( 443 Parse *pParse, /* Parsing context */ 444 SrcList *pTabList, /* The FROM clause */ 445 WhereClause *pWC, /* The WHERE clause */ 446 ExprList *pDistinct /* The result set that needs to be DISTINCT */ 447 ){ 448 Table *pTab; 449 Index *pIdx; 450 int i; 451 int iBase; 452 453 /* If there is more than one table or sub-select in the FROM clause of 454 ** this query, then it will not be possible to show that the DISTINCT 455 ** clause is redundant. */ 456 if( pTabList->nSrc!=1 ) return 0; 457 iBase = pTabList->a[0].iCursor; 458 pTab = pTabList->a[0].pTab; 459 460 /* If any of the expressions is an IPK column on table iBase, then return 461 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 462 ** current SELECT is a correlated sub-query. 463 */ 464 for(i=0; i<pDistinct->nExpr; i++){ 465 Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr); 466 if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1; 467 } 468 469 /* Loop through all indices on the table, checking each to see if it makes 470 ** the DISTINCT qualifier redundant. It does so if: 471 ** 472 ** 1. The index is itself UNIQUE, and 473 ** 474 ** 2. All of the columns in the index are either part of the pDistinct 475 ** list, or else the WHERE clause contains a term of the form "col=X", 476 ** where X is a constant value. The collation sequences of the 477 ** comparison and select-list expressions must match those of the index. 478 ** 479 ** 3. All of those index columns for which the WHERE clause does not 480 ** contain a "col=X" term are subject to a NOT NULL constraint. 481 */ 482 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 483 if( !IsUniqueIndex(pIdx) ) continue; 484 for(i=0; i<pIdx->nKeyCol; i++){ 485 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){ 486 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break; 487 if( indexColumnNotNull(pIdx, i)==0 ) break; 488 } 489 } 490 if( i==pIdx->nKeyCol ){ 491 /* This index implies that the DISTINCT qualifier is redundant. */ 492 return 1; 493 } 494 } 495 496 return 0; 497 } 498 499 500 /* 501 ** Estimate the logarithm of the input value to base 2. 502 */ 503 static LogEst estLog(LogEst N){ 504 return N<=10 ? 0 : sqlite3LogEst(N) - 33; 505 } 506 507 /* 508 ** Convert OP_Column opcodes to OP_Copy in previously generated code. 509 ** 510 ** This routine runs over generated VDBE code and translates OP_Column 511 ** opcodes into OP_Copy when the table is being accessed via co-routine 512 ** instead of via table lookup. 513 ** 514 ** If the bIncrRowid parameter is 0, then any OP_Rowid instructions on 515 ** cursor iTabCur are transformed into OP_Null. Or, if bIncrRowid is non-zero, 516 ** then each OP_Rowid is transformed into an instruction to increment the 517 ** value stored in its output register. 518 */ 519 static void translateColumnToCopy( 520 Parse *pParse, /* Parsing context */ 521 int iStart, /* Translate from this opcode to the end */ 522 int iTabCur, /* OP_Column/OP_Rowid references to this table */ 523 int iRegister, /* The first column is in this register */ 524 int bIncrRowid /* If non-zero, transform OP_rowid to OP_AddImm(1) */ 525 ){ 526 Vdbe *v = pParse->pVdbe; 527 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart); 528 int iEnd = sqlite3VdbeCurrentAddr(v); 529 if( pParse->db->mallocFailed ) return; 530 for(; iStart<iEnd; iStart++, pOp++){ 531 if( pOp->p1!=iTabCur ) continue; 532 if( pOp->opcode==OP_Column ){ 533 pOp->opcode = OP_Copy; 534 pOp->p1 = pOp->p2 + iRegister; 535 pOp->p2 = pOp->p3; 536 pOp->p3 = 0; 537 }else if( pOp->opcode==OP_Rowid ){ 538 if( bIncrRowid ){ 539 /* Increment the value stored in the P2 operand of the OP_Rowid. */ 540 pOp->opcode = OP_AddImm; 541 pOp->p1 = pOp->p2; 542 pOp->p2 = 1; 543 }else{ 544 pOp->opcode = OP_Null; 545 pOp->p1 = 0; 546 pOp->p3 = 0; 547 } 548 } 549 } 550 } 551 552 /* 553 ** Two routines for printing the content of an sqlite3_index_info 554 ** structure. Used for testing and debugging only. If neither 555 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 556 ** are no-ops. 557 */ 558 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) 559 static void TRACE_IDX_INPUTS(sqlite3_index_info *p){ 560 int i; 561 if( !sqlite3WhereTrace ) return; 562 for(i=0; i<p->nConstraint; i++){ 563 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", 564 i, 565 p->aConstraint[i].iColumn, 566 p->aConstraint[i].iTermOffset, 567 p->aConstraint[i].op, 568 p->aConstraint[i].usable); 569 } 570 for(i=0; i<p->nOrderBy; i++){ 571 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 572 i, 573 p->aOrderBy[i].iColumn, 574 p->aOrderBy[i].desc); 575 } 576 } 577 static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){ 578 int i; 579 if( !sqlite3WhereTrace ) return; 580 for(i=0; i<p->nConstraint; i++){ 581 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 582 i, 583 p->aConstraintUsage[i].argvIndex, 584 p->aConstraintUsage[i].omit); 585 } 586 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 587 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 588 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 589 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 590 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); 591 } 592 #else 593 #define TRACE_IDX_INPUTS(A) 594 #define TRACE_IDX_OUTPUTS(A) 595 #endif 596 597 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 598 /* 599 ** Return TRUE if the WHERE clause term pTerm is of a form where it 600 ** could be used with an index to access pSrc, assuming an appropriate 601 ** index existed. 602 */ 603 static int termCanDriveIndex( 604 WhereTerm *pTerm, /* WHERE clause term to check */ 605 struct SrcList_item *pSrc, /* Table we are trying to access */ 606 Bitmask notReady /* Tables in outer loops of the join */ 607 ){ 608 char aff; 609 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 610 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0; 611 if( (pSrc->fg.jointype & JT_LEFT) 612 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 613 && (pTerm->eOperator & WO_IS) 614 ){ 615 /* Cannot use an IS term from the WHERE clause as an index driver for 616 ** the RHS of a LEFT JOIN. Such a term can only be used if it is from 617 ** the ON clause. */ 618 return 0; 619 } 620 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 621 if( pTerm->u.leftColumn<0 ) return 0; 622 aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity; 623 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 624 testcase( pTerm->pExpr->op==TK_IS ); 625 return 1; 626 } 627 #endif 628 629 630 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 631 /* 632 ** Generate code to construct the Index object for an automatic index 633 ** and to set up the WhereLevel object pLevel so that the code generator 634 ** makes use of the automatic index. 635 */ 636 static void constructAutomaticIndex( 637 Parse *pParse, /* The parsing context */ 638 WhereClause *pWC, /* The WHERE clause */ 639 struct SrcList_item *pSrc, /* The FROM clause term to get the next index */ 640 Bitmask notReady, /* Mask of cursors that are not available */ 641 WhereLevel *pLevel /* Write new index here */ 642 ){ 643 int nKeyCol; /* Number of columns in the constructed index */ 644 WhereTerm *pTerm; /* A single term of the WHERE clause */ 645 WhereTerm *pWCEnd; /* End of pWC->a[] */ 646 Index *pIdx; /* Object describing the transient index */ 647 Vdbe *v; /* Prepared statement under construction */ 648 int addrInit; /* Address of the initialization bypass jump */ 649 Table *pTable; /* The table being indexed */ 650 int addrTop; /* Top of the index fill loop */ 651 int regRecord; /* Register holding an index record */ 652 int n; /* Column counter */ 653 int i; /* Loop counter */ 654 int mxBitCol; /* Maximum column in pSrc->colUsed */ 655 CollSeq *pColl; /* Collating sequence to on a column */ 656 WhereLoop *pLoop; /* The Loop object */ 657 char *zNotUsed; /* Extra space on the end of pIdx */ 658 Bitmask idxCols; /* Bitmap of columns used for indexing */ 659 Bitmask extraCols; /* Bitmap of additional columns */ 660 u8 sentWarning = 0; /* True if a warnning has been issued */ 661 Expr *pPartial = 0; /* Partial Index Expression */ 662 int iContinue = 0; /* Jump here to skip excluded rows */ 663 struct SrcList_item *pTabItem; /* FROM clause term being indexed */ 664 int addrCounter = 0; /* Address where integer counter is initialized */ 665 int regBase; /* Array of registers where record is assembled */ 666 667 /* Generate code to skip over the creation and initialization of the 668 ** transient index on 2nd and subsequent iterations of the loop. */ 669 v = pParse->pVdbe; 670 assert( v!=0 ); 671 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 672 673 /* Count the number of columns that will be added to the index 674 ** and used to match WHERE clause constraints */ 675 nKeyCol = 0; 676 pTable = pSrc->pTab; 677 pWCEnd = &pWC->a[pWC->nTerm]; 678 pLoop = pLevel->pWLoop; 679 idxCols = 0; 680 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 681 Expr *pExpr = pTerm->pExpr; 682 assert( !ExprHasProperty(pExpr, EP_FromJoin) /* prereq always non-zero */ 683 || pExpr->iRightJoinTable!=pSrc->iCursor /* for the right-hand */ 684 || pLoop->prereq!=0 ); /* table of a LEFT JOIN */ 685 if( pLoop->prereq==0 686 && (pTerm->wtFlags & TERM_VIRTUAL)==0 687 && !ExprHasProperty(pExpr, EP_FromJoin) 688 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) ){ 689 pPartial = sqlite3ExprAnd(pParse->db, pPartial, 690 sqlite3ExprDup(pParse->db, pExpr, 0)); 691 } 692 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 693 int iCol = pTerm->u.leftColumn; 694 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 695 testcase( iCol==BMS ); 696 testcase( iCol==BMS-1 ); 697 if( !sentWarning ){ 698 sqlite3_log(SQLITE_WARNING_AUTOINDEX, 699 "automatic index on %s(%s)", pTable->zName, 700 pTable->aCol[iCol].zName); 701 sentWarning = 1; 702 } 703 if( (idxCols & cMask)==0 ){ 704 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ 705 goto end_auto_index_create; 706 } 707 pLoop->aLTerm[nKeyCol++] = pTerm; 708 idxCols |= cMask; 709 } 710 } 711 } 712 assert( nKeyCol>0 ); 713 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; 714 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED 715 | WHERE_AUTO_INDEX; 716 717 /* Count the number of additional columns needed to create a 718 ** covering index. A "covering index" is an index that contains all 719 ** columns that are needed by the query. With a covering index, the 720 ** original table never needs to be accessed. Automatic indices must 721 ** be a covering index because the index will not be updated if the 722 ** original table changes and the index and table cannot both be used 723 ** if they go out of sync. 724 */ 725 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); 726 mxBitCol = MIN(BMS-1,pTable->nCol); 727 testcase( pTable->nCol==BMS-1 ); 728 testcase( pTable->nCol==BMS-2 ); 729 for(i=0; i<mxBitCol; i++){ 730 if( extraCols & MASKBIT(i) ) nKeyCol++; 731 } 732 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 733 nKeyCol += pTable->nCol - BMS + 1; 734 } 735 736 /* Construct the Index object to describe this index */ 737 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); 738 if( pIdx==0 ) goto end_auto_index_create; 739 pLoop->u.btree.pIndex = pIdx; 740 pIdx->zName = "auto-index"; 741 pIdx->pTable = pTable; 742 n = 0; 743 idxCols = 0; 744 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 745 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 746 int iCol = pTerm->u.leftColumn; 747 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 748 testcase( iCol==BMS-1 ); 749 testcase( iCol==BMS ); 750 if( (idxCols & cMask)==0 ){ 751 Expr *pX = pTerm->pExpr; 752 idxCols |= cMask; 753 pIdx->aiColumn[n] = pTerm->u.leftColumn; 754 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); 755 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY; 756 n++; 757 } 758 } 759 } 760 assert( (u32)n==pLoop->u.btree.nEq ); 761 762 /* Add additional columns needed to make the automatic index into 763 ** a covering index */ 764 for(i=0; i<mxBitCol; i++){ 765 if( extraCols & MASKBIT(i) ){ 766 pIdx->aiColumn[n] = i; 767 pIdx->azColl[n] = sqlite3StrBINARY; 768 n++; 769 } 770 } 771 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 772 for(i=BMS-1; i<pTable->nCol; i++){ 773 pIdx->aiColumn[n] = i; 774 pIdx->azColl[n] = sqlite3StrBINARY; 775 n++; 776 } 777 } 778 assert( n==nKeyCol ); 779 pIdx->aiColumn[n] = XN_ROWID; 780 pIdx->azColl[n] = sqlite3StrBINARY; 781 782 /* Create the automatic index */ 783 assert( pLevel->iIdxCur>=0 ); 784 pLevel->iIdxCur = pParse->nTab++; 785 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); 786 sqlite3VdbeSetP4KeyInfo(pParse, pIdx); 787 VdbeComment((v, "for %s", pTable->zName)); 788 789 /* Fill the automatic index with content */ 790 sqlite3ExprCachePush(pParse); 791 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom]; 792 if( pTabItem->fg.viaCoroutine ){ 793 int regYield = pTabItem->regReturn; 794 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0); 795 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); 796 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 797 VdbeCoverage(v); 798 VdbeComment((v, "next row of \"%s\"", pTabItem->pTab->zName)); 799 }else{ 800 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); 801 } 802 if( pPartial ){ 803 iContinue = sqlite3VdbeMakeLabel(v); 804 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); 805 pLoop->wsFlags |= WHERE_PARTIALIDX; 806 } 807 regRecord = sqlite3GetTempReg(pParse); 808 regBase = sqlite3GenerateIndexKey( 809 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0 810 ); 811 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 812 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 813 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); 814 if( pTabItem->fg.viaCoroutine ){ 815 sqlite3VdbeChangeP2(v, addrCounter, regBase+n); 816 testcase( pParse->db->mallocFailed ); 817 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur, 818 pTabItem->regResult, 1); 819 sqlite3VdbeGoto(v, addrTop); 820 pTabItem->fg.viaCoroutine = 0; 821 }else{ 822 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); 823 } 824 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 825 sqlite3VdbeJumpHere(v, addrTop); 826 sqlite3ReleaseTempReg(pParse, regRecord); 827 sqlite3ExprCachePop(pParse); 828 829 /* Jump here when skipping the initialization */ 830 sqlite3VdbeJumpHere(v, addrInit); 831 832 end_auto_index_create: 833 sqlite3ExprDelete(pParse->db, pPartial); 834 } 835 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 836 837 #ifndef SQLITE_OMIT_VIRTUALTABLE 838 /* 839 ** Allocate and populate an sqlite3_index_info structure. It is the 840 ** responsibility of the caller to eventually release the structure 841 ** by passing the pointer returned by this function to sqlite3_free(). 842 */ 843 static sqlite3_index_info *allocateIndexInfo( 844 Parse *pParse, 845 WhereClause *pWC, 846 Bitmask mUnusable, /* Ignore terms with these prereqs */ 847 struct SrcList_item *pSrc, 848 ExprList *pOrderBy, 849 u16 *pmNoOmit /* Mask of terms not to omit */ 850 ){ 851 int i, j; 852 int nTerm; 853 struct sqlite3_index_constraint *pIdxCons; 854 struct sqlite3_index_orderby *pIdxOrderBy; 855 struct sqlite3_index_constraint_usage *pUsage; 856 WhereTerm *pTerm; 857 int nOrderBy; 858 sqlite3_index_info *pIdxInfo; 859 u16 mNoOmit = 0; 860 861 /* Count the number of possible WHERE clause constraints referring 862 ** to this virtual table */ 863 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 864 if( pTerm->leftCursor != pSrc->iCursor ) continue; 865 if( pTerm->prereqRight & mUnusable ) continue; 866 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 867 testcase( pTerm->eOperator & WO_IN ); 868 testcase( pTerm->eOperator & WO_ISNULL ); 869 testcase( pTerm->eOperator & WO_IS ); 870 testcase( pTerm->eOperator & WO_ALL ); 871 if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV|WO_IS))==0 ) continue; 872 if( pTerm->wtFlags & TERM_VNULL ) continue; 873 assert( pTerm->u.leftColumn>=(-1) ); 874 nTerm++; 875 } 876 877 /* If the ORDER BY clause contains only columns in the current 878 ** virtual table then allocate space for the aOrderBy part of 879 ** the sqlite3_index_info structure. 880 */ 881 nOrderBy = 0; 882 if( pOrderBy ){ 883 int n = pOrderBy->nExpr; 884 for(i=0; i<n; i++){ 885 Expr *pExpr = pOrderBy->a[i].pExpr; 886 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; 887 } 888 if( i==n){ 889 nOrderBy = n; 890 } 891 } 892 893 /* Allocate the sqlite3_index_info structure 894 */ 895 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 896 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 897 + sizeof(*pIdxOrderBy)*nOrderBy ); 898 if( pIdxInfo==0 ){ 899 sqlite3ErrorMsg(pParse, "out of memory"); 900 return 0; 901 } 902 903 /* Initialize the structure. The sqlite3_index_info structure contains 904 ** many fields that are declared "const" to prevent xBestIndex from 905 ** changing them. We have to do some funky casting in order to 906 ** initialize those fields. 907 */ 908 pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1]; 909 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 910 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 911 *(int*)&pIdxInfo->nConstraint = nTerm; 912 *(int*)&pIdxInfo->nOrderBy = nOrderBy; 913 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons; 914 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy; 915 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage = 916 pUsage; 917 918 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 919 u8 op; 920 if( pTerm->leftCursor != pSrc->iCursor ) continue; 921 if( pTerm->prereqRight & mUnusable ) continue; 922 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 923 testcase( pTerm->eOperator & WO_IN ); 924 testcase( pTerm->eOperator & WO_IS ); 925 testcase( pTerm->eOperator & WO_ISNULL ); 926 testcase( pTerm->eOperator & WO_ALL ); 927 if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV|WO_IS))==0 ) continue; 928 if( pTerm->wtFlags & TERM_VNULL ) continue; 929 assert( pTerm->u.leftColumn>=(-1) ); 930 pIdxCons[j].iColumn = pTerm->u.leftColumn; 931 pIdxCons[j].iTermOffset = i; 932 op = (u8)pTerm->eOperator & WO_ALL; 933 if( op==WO_IN ) op = WO_EQ; 934 if( op==WO_MATCH ){ 935 op = pTerm->eMatchOp; 936 } 937 pIdxCons[j].op = op; 938 /* The direct assignment in the previous line is possible only because 939 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 940 ** following asserts verify this fact. */ 941 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 942 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 943 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 944 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 945 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 946 assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH ); 947 assert( pTerm->eOperator & (WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) ); 948 949 if( op & (WO_LT|WO_LE|WO_GT|WO_GE) 950 && sqlite3ExprIsVector(pTerm->pExpr->pRight) 951 ){ 952 if( i<16 ) mNoOmit |= (1 << i); 953 if( op==WO_LT ) pIdxCons[j].op = WO_LE; 954 if( op==WO_GT ) pIdxCons[j].op = WO_GE; 955 } 956 957 j++; 958 } 959 for(i=0; i<nOrderBy; i++){ 960 Expr *pExpr = pOrderBy->a[i].pExpr; 961 pIdxOrderBy[i].iColumn = pExpr->iColumn; 962 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder; 963 } 964 965 *pmNoOmit = mNoOmit; 966 return pIdxInfo; 967 } 968 969 /* 970 ** The table object reference passed as the second argument to this function 971 ** must represent a virtual table. This function invokes the xBestIndex() 972 ** method of the virtual table with the sqlite3_index_info object that 973 ** comes in as the 3rd argument to this function. 974 ** 975 ** If an error occurs, pParse is populated with an error message and a 976 ** non-zero value is returned. Otherwise, 0 is returned and the output 977 ** part of the sqlite3_index_info structure is left populated. 978 ** 979 ** Whether or not an error is returned, it is the responsibility of the 980 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 981 ** that this is required. 982 */ 983 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 984 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 985 int rc; 986 987 TRACE_IDX_INPUTS(p); 988 rc = pVtab->pModule->xBestIndex(pVtab, p); 989 TRACE_IDX_OUTPUTS(p); 990 991 if( rc!=SQLITE_OK ){ 992 if( rc==SQLITE_NOMEM ){ 993 sqlite3OomFault(pParse->db); 994 }else if( !pVtab->zErrMsg ){ 995 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 996 }else{ 997 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 998 } 999 } 1000 sqlite3_free(pVtab->zErrMsg); 1001 pVtab->zErrMsg = 0; 1002 1003 #if 0 1004 /* This error is now caught by the caller. 1005 ** Search for "xBestIndex malfunction" below */ 1006 for(i=0; i<p->nConstraint; i++){ 1007 if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){ 1008 sqlite3ErrorMsg(pParse, 1009 "table %s: xBestIndex returned an invalid plan", pTab->zName); 1010 } 1011 } 1012 #endif 1013 1014 return pParse->nErr; 1015 } 1016 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 1017 1018 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1019 /* 1020 ** Estimate the location of a particular key among all keys in an 1021 ** index. Store the results in aStat as follows: 1022 ** 1023 ** aStat[0] Est. number of rows less than pRec 1024 ** aStat[1] Est. number of rows equal to pRec 1025 ** 1026 ** Return the index of the sample that is the smallest sample that 1027 ** is greater than or equal to pRec. Note that this index is not an index 1028 ** into the aSample[] array - it is an index into a virtual set of samples 1029 ** based on the contents of aSample[] and the number of fields in record 1030 ** pRec. 1031 */ 1032 static int whereKeyStats( 1033 Parse *pParse, /* Database connection */ 1034 Index *pIdx, /* Index to consider domain of */ 1035 UnpackedRecord *pRec, /* Vector of values to consider */ 1036 int roundUp, /* Round up if true. Round down if false */ 1037 tRowcnt *aStat /* OUT: stats written here */ 1038 ){ 1039 IndexSample *aSample = pIdx->aSample; 1040 int iCol; /* Index of required stats in anEq[] etc. */ 1041 int i; /* Index of first sample >= pRec */ 1042 int iSample; /* Smallest sample larger than or equal to pRec */ 1043 int iMin = 0; /* Smallest sample not yet tested */ 1044 int iTest; /* Next sample to test */ 1045 int res; /* Result of comparison operation */ 1046 int nField; /* Number of fields in pRec */ 1047 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 1048 1049 #ifndef SQLITE_DEBUG 1050 UNUSED_PARAMETER( pParse ); 1051 #endif 1052 assert( pRec!=0 ); 1053 assert( pIdx->nSample>0 ); 1054 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1055 1056 /* Do a binary search to find the first sample greater than or equal 1057 ** to pRec. If pRec contains a single field, the set of samples to search 1058 ** is simply the aSample[] array. If the samples in aSample[] contain more 1059 ** than one fields, all fields following the first are ignored. 1060 ** 1061 ** If pRec contains N fields, where N is more than one, then as well as the 1062 ** samples in aSample[] (truncated to N fields), the search also has to 1063 ** consider prefixes of those samples. For example, if the set of samples 1064 ** in aSample is: 1065 ** 1066 ** aSample[0] = (a, 5) 1067 ** aSample[1] = (a, 10) 1068 ** aSample[2] = (b, 5) 1069 ** aSample[3] = (c, 100) 1070 ** aSample[4] = (c, 105) 1071 ** 1072 ** Then the search space should ideally be the samples above and the 1073 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1074 ** the code actually searches this set: 1075 ** 1076 ** 0: (a) 1077 ** 1: (a, 5) 1078 ** 2: (a, 10) 1079 ** 3: (a, 10) 1080 ** 4: (b) 1081 ** 5: (b, 5) 1082 ** 6: (c) 1083 ** 7: (c, 100) 1084 ** 8: (c, 105) 1085 ** 9: (c, 105) 1086 ** 1087 ** For each sample in the aSample[] array, N samples are present in the 1088 ** effective sample array. In the above, samples 0 and 1 are based on 1089 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1090 ** 1091 ** Often, sample i of each block of N effective samples has (i+1) fields. 1092 ** Except, each sample may be extended to ensure that it is greater than or 1093 ** equal to the previous sample in the array. For example, in the above, 1094 ** sample 2 is the first sample of a block of N samples, so at first it 1095 ** appears that it should be 1 field in size. However, that would make it 1096 ** smaller than sample 1, so the binary search would not work. As a result, 1097 ** it is extended to two fields. The duplicates that this creates do not 1098 ** cause any problems. 1099 */ 1100 nField = pRec->nField; 1101 iCol = 0; 1102 iSample = pIdx->nSample * nField; 1103 do{ 1104 int iSamp; /* Index in aSample[] of test sample */ 1105 int n; /* Number of fields in test sample */ 1106 1107 iTest = (iMin+iSample)/2; 1108 iSamp = iTest / nField; 1109 if( iSamp>0 ){ 1110 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1111 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1112 ** fields that is greater than the previous effective sample. */ 1113 for(n=(iTest % nField) + 1; n<nField; n++){ 1114 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1115 } 1116 }else{ 1117 n = iTest + 1; 1118 } 1119 1120 pRec->nField = n; 1121 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1122 if( res<0 ){ 1123 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1124 iMin = iTest+1; 1125 }else if( res==0 && n<nField ){ 1126 iLower = aSample[iSamp].anLt[n-1]; 1127 iMin = iTest+1; 1128 res = -1; 1129 }else{ 1130 iSample = iTest; 1131 iCol = n-1; 1132 } 1133 }while( res && iMin<iSample ); 1134 i = iSample / nField; 1135 1136 #ifdef SQLITE_DEBUG 1137 /* The following assert statements check that the binary search code 1138 ** above found the right answer. This block serves no purpose other 1139 ** than to invoke the asserts. */ 1140 if( pParse->db->mallocFailed==0 ){ 1141 if( res==0 ){ 1142 /* If (res==0) is true, then pRec must be equal to sample i. */ 1143 assert( i<pIdx->nSample ); 1144 assert( iCol==nField-1 ); 1145 pRec->nField = nField; 1146 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1147 || pParse->db->mallocFailed 1148 ); 1149 }else{ 1150 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1151 ** all samples in the aSample[] array, pRec must be smaller than the 1152 ** (iCol+1) field prefix of sample i. */ 1153 assert( i<=pIdx->nSample && i>=0 ); 1154 pRec->nField = iCol+1; 1155 assert( i==pIdx->nSample 1156 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1157 || pParse->db->mallocFailed ); 1158 1159 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1160 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1161 ** be greater than or equal to the (iCol) field prefix of sample i. 1162 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1163 if( iCol>0 ){ 1164 pRec->nField = iCol; 1165 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1166 || pParse->db->mallocFailed ); 1167 } 1168 if( i>0 ){ 1169 pRec->nField = nField; 1170 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1171 || pParse->db->mallocFailed ); 1172 } 1173 } 1174 } 1175 #endif /* ifdef SQLITE_DEBUG */ 1176 1177 if( res==0 ){ 1178 /* Record pRec is equal to sample i */ 1179 assert( iCol==nField-1 ); 1180 aStat[0] = aSample[i].anLt[iCol]; 1181 aStat[1] = aSample[i].anEq[iCol]; 1182 }else{ 1183 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1184 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1185 ** is larger than all samples in the array. */ 1186 tRowcnt iUpper, iGap; 1187 if( i>=pIdx->nSample ){ 1188 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1189 }else{ 1190 iUpper = aSample[i].anLt[iCol]; 1191 } 1192 1193 if( iLower>=iUpper ){ 1194 iGap = 0; 1195 }else{ 1196 iGap = iUpper - iLower; 1197 } 1198 if( roundUp ){ 1199 iGap = (iGap*2)/3; 1200 }else{ 1201 iGap = iGap/3; 1202 } 1203 aStat[0] = iLower + iGap; 1204 aStat[1] = pIdx->aAvgEq[nField-1]; 1205 } 1206 1207 /* Restore the pRec->nField value before returning. */ 1208 pRec->nField = nField; 1209 return i; 1210 } 1211 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1212 1213 /* 1214 ** If it is not NULL, pTerm is a term that provides an upper or lower 1215 ** bound on a range scan. Without considering pTerm, it is estimated 1216 ** that the scan will visit nNew rows. This function returns the number 1217 ** estimated to be visited after taking pTerm into account. 1218 ** 1219 ** If the user explicitly specified a likelihood() value for this term, 1220 ** then the return value is the likelihood multiplied by the number of 1221 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1222 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1223 */ 1224 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1225 LogEst nRet = nNew; 1226 if( pTerm ){ 1227 if( pTerm->truthProb<=0 ){ 1228 nRet += pTerm->truthProb; 1229 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1230 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1231 } 1232 } 1233 return nRet; 1234 } 1235 1236 1237 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1238 /* 1239 ** Return the affinity for a single column of an index. 1240 */ 1241 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1242 assert( iCol>=0 && iCol<pIdx->nColumn ); 1243 if( !pIdx->zColAff ){ 1244 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1245 } 1246 return pIdx->zColAff[iCol]; 1247 } 1248 #endif 1249 1250 1251 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1252 /* 1253 ** This function is called to estimate the number of rows visited by a 1254 ** range-scan on a skip-scan index. For example: 1255 ** 1256 ** CREATE INDEX i1 ON t1(a, b, c); 1257 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1258 ** 1259 ** Value pLoop->nOut is currently set to the estimated number of rows 1260 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1261 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1262 ** on the stat4 data for the index. this scan will be peformed multiple 1263 ** times (once for each (a,b) combination that matches a=?) is dealt with 1264 ** by the caller. 1265 ** 1266 ** It does this by scanning through all stat4 samples, comparing values 1267 ** extracted from pLower and pUpper with the corresponding column in each 1268 ** sample. If L and U are the number of samples found to be less than or 1269 ** equal to the values extracted from pLower and pUpper respectively, and 1270 ** N is the total number of samples, the pLoop->nOut value is adjusted 1271 ** as follows: 1272 ** 1273 ** nOut = nOut * ( min(U - L, 1) / N ) 1274 ** 1275 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1276 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1277 ** U is set to N. 1278 ** 1279 ** Normally, this function sets *pbDone to 1 before returning. However, 1280 ** if no value can be extracted from either pLower or pUpper (and so the 1281 ** estimate of the number of rows delivered remains unchanged), *pbDone 1282 ** is left as is. 1283 ** 1284 ** If an error occurs, an SQLite error code is returned. Otherwise, 1285 ** SQLITE_OK. 1286 */ 1287 static int whereRangeSkipScanEst( 1288 Parse *pParse, /* Parsing & code generating context */ 1289 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1290 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1291 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1292 int *pbDone /* Set to true if at least one expr. value extracted */ 1293 ){ 1294 Index *p = pLoop->u.btree.pIndex; 1295 int nEq = pLoop->u.btree.nEq; 1296 sqlite3 *db = pParse->db; 1297 int nLower = -1; 1298 int nUpper = p->nSample+1; 1299 int rc = SQLITE_OK; 1300 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1301 CollSeq *pColl; 1302 1303 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1304 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1305 sqlite3_value *pVal = 0; /* Value extracted from record */ 1306 1307 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1308 if( pLower ){ 1309 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1310 nLower = 0; 1311 } 1312 if( pUpper && rc==SQLITE_OK ){ 1313 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1314 nUpper = p2 ? 0 : p->nSample; 1315 } 1316 1317 if( p1 || p2 ){ 1318 int i; 1319 int nDiff; 1320 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1321 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1322 if( rc==SQLITE_OK && p1 ){ 1323 int res = sqlite3MemCompare(p1, pVal, pColl); 1324 if( res>=0 ) nLower++; 1325 } 1326 if( rc==SQLITE_OK && p2 ){ 1327 int res = sqlite3MemCompare(p2, pVal, pColl); 1328 if( res>=0 ) nUpper++; 1329 } 1330 } 1331 nDiff = (nUpper - nLower); 1332 if( nDiff<=0 ) nDiff = 1; 1333 1334 /* If there is both an upper and lower bound specified, and the 1335 ** comparisons indicate that they are close together, use the fallback 1336 ** method (assume that the scan visits 1/64 of the rows) for estimating 1337 ** the number of rows visited. Otherwise, estimate the number of rows 1338 ** using the method described in the header comment for this function. */ 1339 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1340 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1341 pLoop->nOut -= nAdjust; 1342 *pbDone = 1; 1343 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1344 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1345 } 1346 1347 }else{ 1348 assert( *pbDone==0 ); 1349 } 1350 1351 sqlite3ValueFree(p1); 1352 sqlite3ValueFree(p2); 1353 sqlite3ValueFree(pVal); 1354 1355 return rc; 1356 } 1357 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1358 1359 /* 1360 ** This function is used to estimate the number of rows that will be visited 1361 ** by scanning an index for a range of values. The range may have an upper 1362 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1363 ** and lower bounds are represented by pLower and pUpper respectively. For 1364 ** example, assuming that index p is on t1(a): 1365 ** 1366 ** ... FROM t1 WHERE a > ? AND a < ? ... 1367 ** |_____| |_____| 1368 ** | | 1369 ** pLower pUpper 1370 ** 1371 ** If either of the upper or lower bound is not present, then NULL is passed in 1372 ** place of the corresponding WhereTerm. 1373 ** 1374 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1375 ** column subject to the range constraint. Or, equivalently, the number of 1376 ** equality constraints optimized by the proposed index scan. For example, 1377 ** assuming index p is on t1(a, b), and the SQL query is: 1378 ** 1379 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1380 ** 1381 ** then nEq is set to 1 (as the range restricted column, b, is the second 1382 ** left-most column of the index). Or, if the query is: 1383 ** 1384 ** ... FROM t1 WHERE a > ? AND a < ? ... 1385 ** 1386 ** then nEq is set to 0. 1387 ** 1388 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1389 ** number of rows that the index scan is expected to visit without 1390 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1391 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1392 ** to account for the range constraints pLower and pUpper. 1393 ** 1394 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1395 ** used, a single range inequality reduces the search space by a factor of 4. 1396 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1397 ** rows visited by a factor of 64. 1398 */ 1399 static int whereRangeScanEst( 1400 Parse *pParse, /* Parsing & code generating context */ 1401 WhereLoopBuilder *pBuilder, 1402 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1403 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1404 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1405 ){ 1406 int rc = SQLITE_OK; 1407 int nOut = pLoop->nOut; 1408 LogEst nNew; 1409 1410 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1411 Index *p = pLoop->u.btree.pIndex; 1412 int nEq = pLoop->u.btree.nEq; 1413 1414 if( p->nSample>0 && nEq<p->nSampleCol ){ 1415 if( nEq==pBuilder->nRecValid ){ 1416 UnpackedRecord *pRec = pBuilder->pRec; 1417 tRowcnt a[2]; 1418 int nBtm = pLoop->u.btree.nBtm; 1419 int nTop = pLoop->u.btree.nTop; 1420 1421 /* Variable iLower will be set to the estimate of the number of rows in 1422 ** the index that are less than the lower bound of the range query. The 1423 ** lower bound being the concatenation of $P and $L, where $P is the 1424 ** key-prefix formed by the nEq values matched against the nEq left-most 1425 ** columns of the index, and $L is the value in pLower. 1426 ** 1427 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1428 ** is not a simple variable or literal value), the lower bound of the 1429 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1430 ** if $L is available, whereKeyStats() is called for both ($P) and 1431 ** ($P:$L) and the larger of the two returned values is used. 1432 ** 1433 ** Similarly, iUpper is to be set to the estimate of the number of rows 1434 ** less than the upper bound of the range query. Where the upper bound 1435 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1436 ** of iUpper are requested of whereKeyStats() and the smaller used. 1437 ** 1438 ** The number of rows between the two bounds is then just iUpper-iLower. 1439 */ 1440 tRowcnt iLower; /* Rows less than the lower bound */ 1441 tRowcnt iUpper; /* Rows less than the upper bound */ 1442 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1443 int iUprIdx = -1; /* aSample[] for the upper bound */ 1444 1445 if( pRec ){ 1446 testcase( pRec->nField!=pBuilder->nRecValid ); 1447 pRec->nField = pBuilder->nRecValid; 1448 } 1449 /* Determine iLower and iUpper using ($P) only. */ 1450 if( nEq==0 ){ 1451 iLower = 0; 1452 iUpper = p->nRowEst0; 1453 }else{ 1454 /* Note: this call could be optimized away - since the same values must 1455 ** have been requested when testing key $P in whereEqualScanEst(). */ 1456 whereKeyStats(pParse, p, pRec, 0, a); 1457 iLower = a[0]; 1458 iUpper = a[0] + a[1]; 1459 } 1460 1461 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1462 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1463 assert( p->aSortOrder!=0 ); 1464 if( p->aSortOrder[nEq] ){ 1465 /* The roles of pLower and pUpper are swapped for a DESC index */ 1466 SWAP(WhereTerm*, pLower, pUpper); 1467 SWAP(int, nBtm, nTop); 1468 } 1469 1470 /* If possible, improve on the iLower estimate using ($P:$L). */ 1471 if( pLower ){ 1472 int n; /* Values extracted from pExpr */ 1473 Expr *pExpr = pLower->pExpr->pRight; 1474 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1475 if( rc==SQLITE_OK && n ){ 1476 tRowcnt iNew; 1477 u16 mask = WO_GT|WO_LE; 1478 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1479 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1480 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1481 if( iNew>iLower ) iLower = iNew; 1482 nOut--; 1483 pLower = 0; 1484 } 1485 } 1486 1487 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1488 if( pUpper ){ 1489 int n; /* Values extracted from pExpr */ 1490 Expr *pExpr = pUpper->pExpr->pRight; 1491 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1492 if( rc==SQLITE_OK && n ){ 1493 tRowcnt iNew; 1494 u16 mask = WO_GT|WO_LE; 1495 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1496 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1497 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1498 if( iNew<iUpper ) iUpper = iNew; 1499 nOut--; 1500 pUpper = 0; 1501 } 1502 } 1503 1504 pBuilder->pRec = pRec; 1505 if( rc==SQLITE_OK ){ 1506 if( iUpper>iLower ){ 1507 nNew = sqlite3LogEst(iUpper - iLower); 1508 /* TUNING: If both iUpper and iLower are derived from the same 1509 ** sample, then assume they are 4x more selective. This brings 1510 ** the estimated selectivity more in line with what it would be 1511 ** if estimated without the use of STAT3/4 tables. */ 1512 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1513 }else{ 1514 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1515 } 1516 if( nNew<nOut ){ 1517 nOut = nNew; 1518 } 1519 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1520 (u32)iLower, (u32)iUpper, nOut)); 1521 } 1522 }else{ 1523 int bDone = 0; 1524 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1525 if( bDone ) return rc; 1526 } 1527 } 1528 #else 1529 UNUSED_PARAMETER(pParse); 1530 UNUSED_PARAMETER(pBuilder); 1531 assert( pLower || pUpper ); 1532 #endif 1533 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1534 nNew = whereRangeAdjust(pLower, nOut); 1535 nNew = whereRangeAdjust(pUpper, nNew); 1536 1537 /* TUNING: If there is both an upper and lower limit and neither limit 1538 ** has an application-defined likelihood(), assume the range is 1539 ** reduced by an additional 75%. This means that, by default, an open-ended 1540 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1541 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1542 ** match 1/64 of the index. */ 1543 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1544 nNew -= 20; 1545 } 1546 1547 nOut -= (pLower!=0) + (pUpper!=0); 1548 if( nNew<10 ) nNew = 10; 1549 if( nNew<nOut ) nOut = nNew; 1550 #if defined(WHERETRACE_ENABLED) 1551 if( pLoop->nOut>nOut ){ 1552 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1553 pLoop->nOut, nOut)); 1554 } 1555 #endif 1556 pLoop->nOut = (LogEst)nOut; 1557 return rc; 1558 } 1559 1560 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1561 /* 1562 ** Estimate the number of rows that will be returned based on 1563 ** an equality constraint x=VALUE and where that VALUE occurs in 1564 ** the histogram data. This only works when x is the left-most 1565 ** column of an index and sqlite_stat3 histogram data is available 1566 ** for that index. When pExpr==NULL that means the constraint is 1567 ** "x IS NULL" instead of "x=VALUE". 1568 ** 1569 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1570 ** If unable to make an estimate, leave *pnRow unchanged and return 1571 ** non-zero. 1572 ** 1573 ** This routine can fail if it is unable to load a collating sequence 1574 ** required for string comparison, or if unable to allocate memory 1575 ** for a UTF conversion required for comparison. The error is stored 1576 ** in the pParse structure. 1577 */ 1578 static int whereEqualScanEst( 1579 Parse *pParse, /* Parsing & code generating context */ 1580 WhereLoopBuilder *pBuilder, 1581 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1582 tRowcnt *pnRow /* Write the revised row estimate here */ 1583 ){ 1584 Index *p = pBuilder->pNew->u.btree.pIndex; 1585 int nEq = pBuilder->pNew->u.btree.nEq; 1586 UnpackedRecord *pRec = pBuilder->pRec; 1587 int rc; /* Subfunction return code */ 1588 tRowcnt a[2]; /* Statistics */ 1589 int bOk; 1590 1591 assert( nEq>=1 ); 1592 assert( nEq<=p->nColumn ); 1593 assert( p->aSample!=0 ); 1594 assert( p->nSample>0 ); 1595 assert( pBuilder->nRecValid<nEq ); 1596 1597 /* If values are not available for all fields of the index to the left 1598 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1599 if( pBuilder->nRecValid<(nEq-1) ){ 1600 return SQLITE_NOTFOUND; 1601 } 1602 1603 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1604 ** below would return the same value. */ 1605 if( nEq>=p->nColumn ){ 1606 *pnRow = 1; 1607 return SQLITE_OK; 1608 } 1609 1610 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1611 pBuilder->pRec = pRec; 1612 if( rc!=SQLITE_OK ) return rc; 1613 if( bOk==0 ) return SQLITE_NOTFOUND; 1614 pBuilder->nRecValid = nEq; 1615 1616 whereKeyStats(pParse, p, pRec, 0, a); 1617 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1618 p->zName, nEq-1, (int)a[1])); 1619 *pnRow = a[1]; 1620 1621 return rc; 1622 } 1623 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1624 1625 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1626 /* 1627 ** Estimate the number of rows that will be returned based on 1628 ** an IN constraint where the right-hand side of the IN operator 1629 ** is a list of values. Example: 1630 ** 1631 ** WHERE x IN (1,2,3,4) 1632 ** 1633 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1634 ** If unable to make an estimate, leave *pnRow unchanged and return 1635 ** non-zero. 1636 ** 1637 ** This routine can fail if it is unable to load a collating sequence 1638 ** required for string comparison, or if unable to allocate memory 1639 ** for a UTF conversion required for comparison. The error is stored 1640 ** in the pParse structure. 1641 */ 1642 static int whereInScanEst( 1643 Parse *pParse, /* Parsing & code generating context */ 1644 WhereLoopBuilder *pBuilder, 1645 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1646 tRowcnt *pnRow /* Write the revised row estimate here */ 1647 ){ 1648 Index *p = pBuilder->pNew->u.btree.pIndex; 1649 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1650 int nRecValid = pBuilder->nRecValid; 1651 int rc = SQLITE_OK; /* Subfunction return code */ 1652 tRowcnt nEst; /* Number of rows for a single term */ 1653 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1654 int i; /* Loop counter */ 1655 1656 assert( p->aSample!=0 ); 1657 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1658 nEst = nRow0; 1659 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1660 nRowEst += nEst; 1661 pBuilder->nRecValid = nRecValid; 1662 } 1663 1664 if( rc==SQLITE_OK ){ 1665 if( nRowEst > nRow0 ) nRowEst = nRow0; 1666 *pnRow = nRowEst; 1667 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1668 } 1669 assert( pBuilder->nRecValid==nRecValid ); 1670 return rc; 1671 } 1672 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1673 1674 1675 #ifdef WHERETRACE_ENABLED 1676 /* 1677 ** Print the content of a WhereTerm object 1678 */ 1679 static void whereTermPrint(WhereTerm *pTerm, int iTerm){ 1680 if( pTerm==0 ){ 1681 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1682 }else{ 1683 char zType[4]; 1684 char zLeft[50]; 1685 memcpy(zType, "...", 4); 1686 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1687 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1688 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1689 if( pTerm->eOperator & WO_SINGLE ){ 1690 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 1691 pTerm->leftCursor, pTerm->u.leftColumn); 1692 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 1693 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%lld", 1694 pTerm->u.pOrInfo->indexable); 1695 }else{ 1696 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 1697 } 1698 sqlite3DebugPrintf( 1699 "TERM-%-3d %p %s %-12s prob=%-3d op=0x%03x wtFlags=0x%04x", 1700 iTerm, pTerm, zType, zLeft, pTerm->truthProb, 1701 pTerm->eOperator, pTerm->wtFlags); 1702 if( pTerm->iField ){ 1703 sqlite3DebugPrintf(" iField=%d\n", pTerm->iField); 1704 }else{ 1705 sqlite3DebugPrintf("\n"); 1706 } 1707 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 1708 } 1709 } 1710 #endif 1711 1712 #ifdef WHERETRACE_ENABLED 1713 /* 1714 ** Show the complete content of a WhereClause 1715 */ 1716 void sqlite3WhereClausePrint(WhereClause *pWC){ 1717 int i; 1718 for(i=0; i<pWC->nTerm; i++){ 1719 whereTermPrint(&pWC->a[i], i); 1720 } 1721 } 1722 #endif 1723 1724 #ifdef WHERETRACE_ENABLED 1725 /* 1726 ** Print a WhereLoop object for debugging purposes 1727 */ 1728 static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){ 1729 WhereInfo *pWInfo = pWC->pWInfo; 1730 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 1731 struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab; 1732 Table *pTab = pItem->pTab; 1733 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 1734 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 1735 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 1736 sqlite3DebugPrintf(" %12s", 1737 pItem->zAlias ? pItem->zAlias : pTab->zName); 1738 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 1739 const char *zName; 1740 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 1741 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 1742 int i = sqlite3Strlen30(zName) - 1; 1743 while( zName[i]!='_' ) i--; 1744 zName += i; 1745 } 1746 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 1747 }else{ 1748 sqlite3DebugPrintf("%20s",""); 1749 } 1750 }else{ 1751 char *z; 1752 if( p->u.vtab.idxStr ){ 1753 z = sqlite3_mprintf("(%d,\"%s\",%x)", 1754 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 1755 }else{ 1756 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 1757 } 1758 sqlite3DebugPrintf(" %-19s", z); 1759 sqlite3_free(z); 1760 } 1761 if( p->wsFlags & WHERE_SKIPSCAN ){ 1762 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 1763 }else{ 1764 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); 1765 } 1766 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 1767 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 1768 int i; 1769 for(i=0; i<p->nLTerm; i++){ 1770 whereTermPrint(p->aLTerm[i], i); 1771 } 1772 } 1773 } 1774 #endif 1775 1776 /* 1777 ** Convert bulk memory into a valid WhereLoop that can be passed 1778 ** to whereLoopClear harmlessly. 1779 */ 1780 static void whereLoopInit(WhereLoop *p){ 1781 p->aLTerm = p->aLTermSpace; 1782 p->nLTerm = 0; 1783 p->nLSlot = ArraySize(p->aLTermSpace); 1784 p->wsFlags = 0; 1785 } 1786 1787 /* 1788 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 1789 */ 1790 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 1791 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 1792 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 1793 sqlite3_free(p->u.vtab.idxStr); 1794 p->u.vtab.needFree = 0; 1795 p->u.vtab.idxStr = 0; 1796 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 1797 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 1798 sqlite3DbFreeNN(db, p->u.btree.pIndex); 1799 p->u.btree.pIndex = 0; 1800 } 1801 } 1802 } 1803 1804 /* 1805 ** Deallocate internal memory used by a WhereLoop object 1806 */ 1807 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 1808 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1809 whereLoopClearUnion(db, p); 1810 whereLoopInit(p); 1811 } 1812 1813 /* 1814 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 1815 */ 1816 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 1817 WhereTerm **paNew; 1818 if( p->nLSlot>=n ) return SQLITE_OK; 1819 n = (n+7)&~7; 1820 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 1821 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 1822 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 1823 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1824 p->aLTerm = paNew; 1825 p->nLSlot = n; 1826 return SQLITE_OK; 1827 } 1828 1829 /* 1830 ** Transfer content from the second pLoop into the first. 1831 */ 1832 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 1833 whereLoopClearUnion(db, pTo); 1834 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 1835 memset(&pTo->u, 0, sizeof(pTo->u)); 1836 return SQLITE_NOMEM_BKPT; 1837 } 1838 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 1839 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 1840 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 1841 pFrom->u.vtab.needFree = 0; 1842 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 1843 pFrom->u.btree.pIndex = 0; 1844 } 1845 return SQLITE_OK; 1846 } 1847 1848 /* 1849 ** Delete a WhereLoop object 1850 */ 1851 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 1852 whereLoopClear(db, p); 1853 sqlite3DbFreeNN(db, p); 1854 } 1855 1856 /* 1857 ** Free a WhereInfo structure 1858 */ 1859 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 1860 if( ALWAYS(pWInfo) ){ 1861 int i; 1862 for(i=0; i<pWInfo->nLevel; i++){ 1863 WhereLevel *pLevel = &pWInfo->a[i]; 1864 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){ 1865 sqlite3DbFree(db, pLevel->u.in.aInLoop); 1866 } 1867 } 1868 sqlite3WhereClauseClear(&pWInfo->sWC); 1869 while( pWInfo->pLoops ){ 1870 WhereLoop *p = pWInfo->pLoops; 1871 pWInfo->pLoops = p->pNextLoop; 1872 whereLoopDelete(db, p); 1873 } 1874 sqlite3DbFreeNN(db, pWInfo); 1875 } 1876 } 1877 1878 /* 1879 ** Return TRUE if all of the following are true: 1880 ** 1881 ** (1) X has the same or lower cost that Y 1882 ** (2) X is a proper subset of Y 1883 ** (3) X skips at least as many columns as Y 1884 ** 1885 ** By "proper subset" we mean that X uses fewer WHERE clause terms 1886 ** than Y and that every WHERE clause term used by X is also used 1887 ** by Y. 1888 ** 1889 ** If X is a proper subset of Y then Y is a better choice and ought 1890 ** to have a lower cost. This routine returns TRUE when that cost 1891 ** relationship is inverted and needs to be adjusted. The third rule 1892 ** was added because if X uses skip-scan less than Y it still might 1893 ** deserve a lower cost even if it is a proper subset of Y. 1894 */ 1895 static int whereLoopCheaperProperSubset( 1896 const WhereLoop *pX, /* First WhereLoop to compare */ 1897 const WhereLoop *pY /* Compare against this WhereLoop */ 1898 ){ 1899 int i, j; 1900 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 1901 return 0; /* X is not a subset of Y */ 1902 } 1903 if( pY->nSkip > pX->nSkip ) return 0; 1904 if( pX->rRun >= pY->rRun ){ 1905 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ 1906 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ 1907 } 1908 for(i=pX->nLTerm-1; i>=0; i--){ 1909 if( pX->aLTerm[i]==0 ) continue; 1910 for(j=pY->nLTerm-1; j>=0; j--){ 1911 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 1912 } 1913 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 1914 } 1915 return 1; /* All conditions meet */ 1916 } 1917 1918 /* 1919 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so 1920 ** that: 1921 ** 1922 ** (1) pTemplate costs less than any other WhereLoops that are a proper 1923 ** subset of pTemplate 1924 ** 1925 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 1926 ** is a proper subset. 1927 ** 1928 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 1929 ** WHERE clause terms than Y and that every WHERE clause term used by X is 1930 ** also used by Y. 1931 */ 1932 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 1933 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 1934 for(; p; p=p->pNextLoop){ 1935 if( p->iTab!=pTemplate->iTab ) continue; 1936 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 1937 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 1938 /* Adjust pTemplate cost downward so that it is cheaper than its 1939 ** subset p. */ 1940 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 1941 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); 1942 pTemplate->rRun = p->rRun; 1943 pTemplate->nOut = p->nOut - 1; 1944 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 1945 /* Adjust pTemplate cost upward so that it is costlier than p since 1946 ** pTemplate is a proper subset of p */ 1947 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 1948 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); 1949 pTemplate->rRun = p->rRun; 1950 pTemplate->nOut = p->nOut + 1; 1951 } 1952 } 1953 } 1954 1955 /* 1956 ** Search the list of WhereLoops in *ppPrev looking for one that can be 1957 ** replaced by pTemplate. 1958 ** 1959 ** Return NULL if pTemplate does not belong on the WhereLoop list. 1960 ** In other words if pTemplate ought to be dropped from further consideration. 1961 ** 1962 ** If pX is a WhereLoop that pTemplate can replace, then return the 1963 ** link that points to pX. 1964 ** 1965 ** If pTemplate cannot replace any existing element of the list but needs 1966 ** to be added to the list as a new entry, then return a pointer to the 1967 ** tail of the list. 1968 */ 1969 static WhereLoop **whereLoopFindLesser( 1970 WhereLoop **ppPrev, 1971 const WhereLoop *pTemplate 1972 ){ 1973 WhereLoop *p; 1974 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 1975 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 1976 /* If either the iTab or iSortIdx values for two WhereLoop are different 1977 ** then those WhereLoops need to be considered separately. Neither is 1978 ** a candidate to replace the other. */ 1979 continue; 1980 } 1981 /* In the current implementation, the rSetup value is either zero 1982 ** or the cost of building an automatic index (NlogN) and the NlogN 1983 ** is the same for compatible WhereLoops. */ 1984 assert( p->rSetup==0 || pTemplate->rSetup==0 1985 || p->rSetup==pTemplate->rSetup ); 1986 1987 /* whereLoopAddBtree() always generates and inserts the automatic index 1988 ** case first. Hence compatible candidate WhereLoops never have a larger 1989 ** rSetup. Call this SETUP-INVARIANT */ 1990 assert( p->rSetup>=pTemplate->rSetup ); 1991 1992 /* Any loop using an appliation-defined index (or PRIMARY KEY or 1993 ** UNIQUE constraint) with one or more == constraints is better 1994 ** than an automatic index. Unless it is a skip-scan. */ 1995 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 1996 && (pTemplate->nSkip)==0 1997 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 1998 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 1999 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2000 ){ 2001 break; 2002 } 2003 2004 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2005 ** discarded. WhereLoop p is better if: 2006 ** (1) p has no more dependencies than pTemplate, and 2007 ** (2) p has an equal or lower cost than pTemplate 2008 */ 2009 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2010 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2011 && p->rRun<=pTemplate->rRun /* (2b) */ 2012 && p->nOut<=pTemplate->nOut /* (2c) */ 2013 ){ 2014 return 0; /* Discard pTemplate */ 2015 } 2016 2017 /* If pTemplate is always better than p, then cause p to be overwritten 2018 ** with pTemplate. pTemplate is better than p if: 2019 ** (1) pTemplate has no more dependences than p, and 2020 ** (2) pTemplate has an equal or lower cost than p. 2021 */ 2022 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2023 && p->rRun>=pTemplate->rRun /* (2a) */ 2024 && p->nOut>=pTemplate->nOut /* (2b) */ 2025 ){ 2026 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2027 break; /* Cause p to be overwritten by pTemplate */ 2028 } 2029 } 2030 return ppPrev; 2031 } 2032 2033 /* 2034 ** Insert or replace a WhereLoop entry using the template supplied. 2035 ** 2036 ** An existing WhereLoop entry might be overwritten if the new template 2037 ** is better and has fewer dependencies. Or the template will be ignored 2038 ** and no insert will occur if an existing WhereLoop is faster and has 2039 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2040 ** added based on the template. 2041 ** 2042 ** If pBuilder->pOrSet is not NULL then we care about only the 2043 ** prerequisites and rRun and nOut costs of the N best loops. That 2044 ** information is gathered in the pBuilder->pOrSet object. This special 2045 ** processing mode is used only for OR clause processing. 2046 ** 2047 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2048 ** still might overwrite similar loops with the new template if the 2049 ** new template is better. Loops may be overwritten if the following 2050 ** conditions are met: 2051 ** 2052 ** (1) They have the same iTab. 2053 ** (2) They have the same iSortIdx. 2054 ** (3) The template has same or fewer dependencies than the current loop 2055 ** (4) The template has the same or lower cost than the current loop 2056 */ 2057 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2058 WhereLoop **ppPrev, *p; 2059 WhereInfo *pWInfo = pBuilder->pWInfo; 2060 sqlite3 *db = pWInfo->pParse->db; 2061 int rc; 2062 2063 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2064 ** and prereqs. 2065 */ 2066 if( pBuilder->pOrSet!=0 ){ 2067 if( pTemplate->nLTerm ){ 2068 #if WHERETRACE_ENABLED 2069 u16 n = pBuilder->pOrSet->n; 2070 int x = 2071 #endif 2072 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2073 pTemplate->nOut); 2074 #if WHERETRACE_ENABLED /* 0x8 */ 2075 if( sqlite3WhereTrace & 0x8 ){ 2076 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2077 whereLoopPrint(pTemplate, pBuilder->pWC); 2078 } 2079 #endif 2080 } 2081 return SQLITE_OK; 2082 } 2083 2084 /* Look for an existing WhereLoop to replace with pTemplate 2085 */ 2086 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2087 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2088 2089 if( ppPrev==0 ){ 2090 /* There already exists a WhereLoop on the list that is better 2091 ** than pTemplate, so just ignore pTemplate */ 2092 #if WHERETRACE_ENABLED /* 0x8 */ 2093 if( sqlite3WhereTrace & 0x8 ){ 2094 sqlite3DebugPrintf(" skip: "); 2095 whereLoopPrint(pTemplate, pBuilder->pWC); 2096 } 2097 #endif 2098 return SQLITE_OK; 2099 }else{ 2100 p = *ppPrev; 2101 } 2102 2103 /* If we reach this point it means that either p[] should be overwritten 2104 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2105 ** WhereLoop and insert it. 2106 */ 2107 #if WHERETRACE_ENABLED /* 0x8 */ 2108 if( sqlite3WhereTrace & 0x8 ){ 2109 if( p!=0 ){ 2110 sqlite3DebugPrintf("replace: "); 2111 whereLoopPrint(p, pBuilder->pWC); 2112 sqlite3DebugPrintf(" with: "); 2113 }else{ 2114 sqlite3DebugPrintf(" add: "); 2115 } 2116 whereLoopPrint(pTemplate, pBuilder->pWC); 2117 } 2118 #endif 2119 if( p==0 ){ 2120 /* Allocate a new WhereLoop to add to the end of the list */ 2121 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2122 if( p==0 ) return SQLITE_NOMEM_BKPT; 2123 whereLoopInit(p); 2124 p->pNextLoop = 0; 2125 }else{ 2126 /* We will be overwriting WhereLoop p[]. But before we do, first 2127 ** go through the rest of the list and delete any other entries besides 2128 ** p[] that are also supplated by pTemplate */ 2129 WhereLoop **ppTail = &p->pNextLoop; 2130 WhereLoop *pToDel; 2131 while( *ppTail ){ 2132 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2133 if( ppTail==0 ) break; 2134 pToDel = *ppTail; 2135 if( pToDel==0 ) break; 2136 *ppTail = pToDel->pNextLoop; 2137 #if WHERETRACE_ENABLED /* 0x8 */ 2138 if( sqlite3WhereTrace & 0x8 ){ 2139 sqlite3DebugPrintf(" delete: "); 2140 whereLoopPrint(pToDel, pBuilder->pWC); 2141 } 2142 #endif 2143 whereLoopDelete(db, pToDel); 2144 } 2145 } 2146 rc = whereLoopXfer(db, p, pTemplate); 2147 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2148 Index *pIndex = p->u.btree.pIndex; 2149 if( pIndex && pIndex->tnum==0 ){ 2150 p->u.btree.pIndex = 0; 2151 } 2152 } 2153 return rc; 2154 } 2155 2156 /* 2157 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2158 ** WHERE clause that reference the loop but which are not used by an 2159 ** index. 2160 * 2161 ** For every WHERE clause term that is not used by the index 2162 ** and which has a truth probability assigned by one of the likelihood(), 2163 ** likely(), or unlikely() SQL functions, reduce the estimated number 2164 ** of output rows by the probability specified. 2165 ** 2166 ** TUNING: For every WHERE clause term that is not used by the index 2167 ** and which does not have an assigned truth probability, heuristics 2168 ** described below are used to try to estimate the truth probability. 2169 ** TODO --> Perhaps this is something that could be improved by better 2170 ** table statistics. 2171 ** 2172 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2173 ** value corresponds to -1 in LogEst notation, so this means decrement 2174 ** the WhereLoop.nOut field for every such WHERE clause term. 2175 ** 2176 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2177 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2178 ** final output row estimate is no greater than 1/4 of the total number 2179 ** of rows in the table. In other words, assume that x==EXPR will filter 2180 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2181 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2182 ** on the "x" column and so in that case only cap the output row estimate 2183 ** at 1/2 instead of 1/4. 2184 */ 2185 static void whereLoopOutputAdjust( 2186 WhereClause *pWC, /* The WHERE clause */ 2187 WhereLoop *pLoop, /* The loop to adjust downward */ 2188 LogEst nRow /* Number of rows in the entire table */ 2189 ){ 2190 WhereTerm *pTerm, *pX; 2191 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2192 int i, j, k; 2193 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2194 2195 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2196 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ 2197 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; 2198 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2199 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2200 for(j=pLoop->nLTerm-1; j>=0; j--){ 2201 pX = pLoop->aLTerm[j]; 2202 if( pX==0 ) continue; 2203 if( pX==pTerm ) break; 2204 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2205 } 2206 if( j<0 ){ 2207 if( pTerm->truthProb<=0 ){ 2208 /* If a truth probability is specified using the likelihood() hints, 2209 ** then use the probability provided by the application. */ 2210 pLoop->nOut += pTerm->truthProb; 2211 }else{ 2212 /* In the absence of explicit truth probabilities, use heuristics to 2213 ** guess a reasonable truth probability. */ 2214 pLoop->nOut--; 2215 if( pTerm->eOperator&(WO_EQ|WO_IS) ){ 2216 Expr *pRight = pTerm->pExpr->pRight; 2217 testcase( pTerm->pExpr->op==TK_IS ); 2218 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2219 k = 10; 2220 }else{ 2221 k = 20; 2222 } 2223 if( iReduce<k ) iReduce = k; 2224 } 2225 } 2226 } 2227 } 2228 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce; 2229 } 2230 2231 /* 2232 ** Term pTerm is a vector range comparison operation. The first comparison 2233 ** in the vector can be optimized using column nEq of the index. This 2234 ** function returns the total number of vector elements that can be used 2235 ** as part of the range comparison. 2236 ** 2237 ** For example, if the query is: 2238 ** 2239 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2240 ** 2241 ** and the index: 2242 ** 2243 ** CREATE INDEX ... ON (a, b, c, d, e) 2244 ** 2245 ** then this function would be invoked with nEq=1. The value returned in 2246 ** this case is 3. 2247 */ 2248 static int whereRangeVectorLen( 2249 Parse *pParse, /* Parsing context */ 2250 int iCur, /* Cursor open on pIdx */ 2251 Index *pIdx, /* The index to be used for a inequality constraint */ 2252 int nEq, /* Number of prior equality constraints on same index */ 2253 WhereTerm *pTerm /* The vector inequality constraint */ 2254 ){ 2255 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2256 int i; 2257 2258 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2259 for(i=1; i<nCmp; i++){ 2260 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2261 ** of the index. If not, exit the loop. */ 2262 char aff; /* Comparison affinity */ 2263 char idxaff = 0; /* Indexed columns affinity */ 2264 CollSeq *pColl; /* Comparison collation sequence */ 2265 Expr *pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2266 Expr *pRhs = pTerm->pExpr->pRight; 2267 if( pRhs->flags & EP_xIsSelect ){ 2268 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2269 }else{ 2270 pRhs = pRhs->x.pList->a[i].pExpr; 2271 } 2272 2273 /* Check that the LHS of the comparison is a column reference to 2274 ** the right column of the right source table. And that the sort 2275 ** order of the index column is the same as the sort order of the 2276 ** leftmost index column. */ 2277 if( pLhs->op!=TK_COLUMN 2278 || pLhs->iTable!=iCur 2279 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2280 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2281 ){ 2282 break; 2283 } 2284 2285 testcase( pLhs->iColumn==XN_ROWID ); 2286 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2287 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2288 if( aff!=idxaff ) break; 2289 2290 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2291 if( pColl==0 ) break; 2292 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2293 } 2294 return i; 2295 } 2296 2297 /* 2298 ** Adjust the cost C by the costMult facter T. This only occurs if 2299 ** compiled with -DSQLITE_ENABLE_COSTMULT 2300 */ 2301 #ifdef SQLITE_ENABLE_COSTMULT 2302 # define ApplyCostMultiplier(C,T) C += T 2303 #else 2304 # define ApplyCostMultiplier(C,T) 2305 #endif 2306 2307 /* 2308 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2309 ** index pIndex. Try to match one more. 2310 ** 2311 ** When this function is called, pBuilder->pNew->nOut contains the 2312 ** number of rows expected to be visited by filtering using the nEq 2313 ** terms only. If it is modified, this value is restored before this 2314 ** function returns. 2315 ** 2316 ** If pProbe->tnum==0, that means pIndex is a fake index used for the 2317 ** INTEGER PRIMARY KEY. 2318 */ 2319 static int whereLoopAddBtreeIndex( 2320 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2321 struct SrcList_item *pSrc, /* FROM clause term being analyzed */ 2322 Index *pProbe, /* An index on pSrc */ 2323 LogEst nInMul /* log(Number of iterations due to IN) */ 2324 ){ 2325 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2326 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2327 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2328 WhereLoop *pNew; /* Template WhereLoop under construction */ 2329 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2330 int opMask; /* Valid operators for constraints */ 2331 WhereScan scan; /* Iterator for WHERE terms */ 2332 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2333 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2334 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2335 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2336 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2337 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2338 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2339 LogEst saved_nOut; /* Original value of pNew->nOut */ 2340 int rc = SQLITE_OK; /* Return code */ 2341 LogEst rSize; /* Number of rows in the table */ 2342 LogEst rLogSize; /* Logarithm of table size */ 2343 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2344 2345 pNew = pBuilder->pNew; 2346 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2347 WHERETRACE(0x800, ("BEGIN addBtreeIdx(%s), nEq=%d\n", 2348 pProbe->zName, pNew->u.btree.nEq)); 2349 2350 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2351 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2352 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2353 opMask = WO_LT|WO_LE; 2354 }else{ 2355 assert( pNew->u.btree.nBtm==0 ); 2356 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2357 } 2358 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2359 2360 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2361 2362 saved_nEq = pNew->u.btree.nEq; 2363 saved_nBtm = pNew->u.btree.nBtm; 2364 saved_nTop = pNew->u.btree.nTop; 2365 saved_nSkip = pNew->nSkip; 2366 saved_nLTerm = pNew->nLTerm; 2367 saved_wsFlags = pNew->wsFlags; 2368 saved_prereq = pNew->prereq; 2369 saved_nOut = pNew->nOut; 2370 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2371 opMask, pProbe); 2372 pNew->rSetup = 0; 2373 rSize = pProbe->aiRowLogEst[0]; 2374 rLogSize = estLog(rSize); 2375 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2376 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2377 LogEst rCostIdx; 2378 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2379 int nIn = 0; 2380 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2381 int nRecValid = pBuilder->nRecValid; 2382 #endif 2383 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2384 && indexColumnNotNull(pProbe, saved_nEq) 2385 ){ 2386 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2387 } 2388 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2389 2390 /* Do not allow the upper bound of a LIKE optimization range constraint 2391 ** to mix with a lower range bound from some other source */ 2392 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2393 2394 /* Do not allow IS constraints from the WHERE clause to be used by the 2395 ** right table of a LEFT JOIN. Only constraints in the ON clause are 2396 ** allowed */ 2397 if( (pSrc->fg.jointype & JT_LEFT)!=0 2398 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2399 && (eOp & (WO_IS|WO_ISNULL))!=0 2400 ){ 2401 testcase( eOp & WO_IS ); 2402 testcase( eOp & WO_ISNULL ); 2403 continue; 2404 } 2405 2406 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2407 pBuilder->bldFlags |= SQLITE_BLDF_UNIQUE; 2408 }else{ 2409 pBuilder->bldFlags |= SQLITE_BLDF_INDEXED; 2410 } 2411 pNew->wsFlags = saved_wsFlags; 2412 pNew->u.btree.nEq = saved_nEq; 2413 pNew->u.btree.nBtm = saved_nBtm; 2414 pNew->u.btree.nTop = saved_nTop; 2415 pNew->nLTerm = saved_nLTerm; 2416 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2417 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2418 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2419 2420 assert( nInMul==0 2421 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2422 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2423 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2424 ); 2425 2426 if( eOp & WO_IN ){ 2427 Expr *pExpr = pTerm->pExpr; 2428 pNew->wsFlags |= WHERE_COLUMN_IN; 2429 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 2430 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2431 int i; 2432 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2433 2434 /* The expression may actually be of the form (x, y) IN (SELECT...). 2435 ** In this case there is a separate term for each of (x) and (y). 2436 ** However, the nIn multiplier should only be applied once, not once 2437 ** for each such term. The following loop checks that pTerm is the 2438 ** first such term in use, and sets nIn back to 0 if it is not. */ 2439 for(i=0; i<pNew->nLTerm-1; i++){ 2440 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2441 } 2442 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2443 /* "x IN (value, value, ...)" */ 2444 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2445 assert( nIn>0 ); /* RHS always has 2 or more terms... The parser 2446 ** changes "x IN (?)" into "x=?". */ 2447 } 2448 }else if( eOp & (WO_EQ|WO_IS) ){ 2449 int iCol = pProbe->aiColumn[saved_nEq]; 2450 pNew->wsFlags |= WHERE_COLUMN_EQ; 2451 assert( saved_nEq==pNew->u.btree.nEq ); 2452 if( iCol==XN_ROWID 2453 || (iCol>0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2454 ){ 2455 if( iCol>=0 && pProbe->uniqNotNull==0 ){ 2456 pNew->wsFlags |= WHERE_UNQ_WANTED; 2457 }else{ 2458 pNew->wsFlags |= WHERE_ONEROW; 2459 } 2460 } 2461 }else if( eOp & WO_ISNULL ){ 2462 pNew->wsFlags |= WHERE_COLUMN_NULL; 2463 }else if( eOp & (WO_GT|WO_GE) ){ 2464 testcase( eOp & WO_GT ); 2465 testcase( eOp & WO_GE ); 2466 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2467 pNew->u.btree.nBtm = whereRangeVectorLen( 2468 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2469 ); 2470 pBtm = pTerm; 2471 pTop = 0; 2472 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2473 /* Range contraints that come from the LIKE optimization are 2474 ** always used in pairs. */ 2475 pTop = &pTerm[1]; 2476 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2477 assert( pTop->wtFlags & TERM_LIKEOPT ); 2478 assert( pTop->eOperator==WO_LT ); 2479 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2480 pNew->aLTerm[pNew->nLTerm++] = pTop; 2481 pNew->wsFlags |= WHERE_TOP_LIMIT; 2482 pNew->u.btree.nTop = 1; 2483 } 2484 }else{ 2485 assert( eOp & (WO_LT|WO_LE) ); 2486 testcase( eOp & WO_LT ); 2487 testcase( eOp & WO_LE ); 2488 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2489 pNew->u.btree.nTop = whereRangeVectorLen( 2490 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2491 ); 2492 pTop = pTerm; 2493 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2494 pNew->aLTerm[pNew->nLTerm-2] : 0; 2495 } 2496 2497 /* At this point pNew->nOut is set to the number of rows expected to 2498 ** be visited by the index scan before considering term pTerm, or the 2499 ** values of nIn and nInMul. In other words, assuming that all 2500 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2501 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2502 assert( pNew->nOut==saved_nOut ); 2503 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2504 /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4 2505 ** data, using some other estimate. */ 2506 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2507 }else{ 2508 int nEq = ++pNew->u.btree.nEq; 2509 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2510 2511 assert( pNew->nOut==saved_nOut ); 2512 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2513 assert( (eOp & WO_IN) || nIn==0 ); 2514 testcase( eOp & WO_IN ); 2515 pNew->nOut += pTerm->truthProb; 2516 pNew->nOut -= nIn; 2517 }else{ 2518 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2519 tRowcnt nOut = 0; 2520 if( nInMul==0 2521 && pProbe->nSample 2522 && pNew->u.btree.nEq<=pProbe->nSampleCol 2523 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) 2524 ){ 2525 Expr *pExpr = pTerm->pExpr; 2526 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2527 testcase( eOp & WO_EQ ); 2528 testcase( eOp & WO_IS ); 2529 testcase( eOp & WO_ISNULL ); 2530 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2531 }else{ 2532 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2533 } 2534 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2535 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2536 if( nOut ){ 2537 pNew->nOut = sqlite3LogEst(nOut); 2538 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2539 pNew->nOut -= nIn; 2540 } 2541 } 2542 if( nOut==0 ) 2543 #endif 2544 { 2545 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2546 if( eOp & WO_ISNULL ){ 2547 /* TUNING: If there is no likelihood() value, assume that a 2548 ** "col IS NULL" expression matches twice as many rows 2549 ** as (col=?). */ 2550 pNew->nOut += 10; 2551 } 2552 } 2553 } 2554 } 2555 2556 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2557 ** it to pNew->rRun, which is currently set to the cost of the index 2558 ** seek only. Then, if this is a non-covering index, add the cost of 2559 ** visiting the rows in the main table. */ 2560 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2561 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2562 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2563 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2564 } 2565 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2566 2567 nOutUnadjusted = pNew->nOut; 2568 pNew->rRun += nInMul + nIn; 2569 pNew->nOut += nInMul + nIn; 2570 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 2571 rc = whereLoopInsert(pBuilder, pNew); 2572 2573 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2574 pNew->nOut = saved_nOut; 2575 }else{ 2576 pNew->nOut = nOutUnadjusted; 2577 } 2578 2579 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 2580 && pNew->u.btree.nEq<pProbe->nColumn 2581 ){ 2582 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 2583 } 2584 pNew->nOut = saved_nOut; 2585 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2586 pBuilder->nRecValid = nRecValid; 2587 #endif 2588 } 2589 pNew->prereq = saved_prereq; 2590 pNew->u.btree.nEq = saved_nEq; 2591 pNew->u.btree.nBtm = saved_nBtm; 2592 pNew->u.btree.nTop = saved_nTop; 2593 pNew->nSkip = saved_nSkip; 2594 pNew->wsFlags = saved_wsFlags; 2595 pNew->nOut = saved_nOut; 2596 pNew->nLTerm = saved_nLTerm; 2597 2598 /* Consider using a skip-scan if there are no WHERE clause constraints 2599 ** available for the left-most terms of the index, and if the average 2600 ** number of repeats in the left-most terms is at least 18. 2601 ** 2602 ** The magic number 18 is selected on the basis that scanning 17 rows 2603 ** is almost always quicker than an index seek (even though if the index 2604 ** contains fewer than 2^17 rows we assume otherwise in other parts of 2605 ** the code). And, even if it is not, it should not be too much slower. 2606 ** On the other hand, the extra seeks could end up being significantly 2607 ** more expensive. */ 2608 assert( 42==sqlite3LogEst(18) ); 2609 if( saved_nEq==saved_nSkip 2610 && saved_nEq+1<pProbe->nKeyCol 2611 && pProbe->noSkipScan==0 2612 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 2613 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 2614 ){ 2615 LogEst nIter; 2616 pNew->u.btree.nEq++; 2617 pNew->nSkip++; 2618 pNew->aLTerm[pNew->nLTerm++] = 0; 2619 pNew->wsFlags |= WHERE_SKIPSCAN; 2620 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 2621 pNew->nOut -= nIter; 2622 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 2623 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 2624 nIter += 5; 2625 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 2626 pNew->nOut = saved_nOut; 2627 pNew->u.btree.nEq = saved_nEq; 2628 pNew->nSkip = saved_nSkip; 2629 pNew->wsFlags = saved_wsFlags; 2630 } 2631 2632 WHERETRACE(0x800, ("END addBtreeIdx(%s), nEq=%d, rc=%d\n", 2633 pProbe->zName, saved_nEq, rc)); 2634 return rc; 2635 } 2636 2637 /* 2638 ** Return True if it is possible that pIndex might be useful in 2639 ** implementing the ORDER BY clause in pBuilder. 2640 ** 2641 ** Return False if pBuilder does not contain an ORDER BY clause or 2642 ** if there is no way for pIndex to be useful in implementing that 2643 ** ORDER BY clause. 2644 */ 2645 static int indexMightHelpWithOrderBy( 2646 WhereLoopBuilder *pBuilder, 2647 Index *pIndex, 2648 int iCursor 2649 ){ 2650 ExprList *pOB; 2651 ExprList *aColExpr; 2652 int ii, jj; 2653 2654 if( pIndex->bUnordered ) return 0; 2655 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 2656 for(ii=0; ii<pOB->nExpr; ii++){ 2657 Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr); 2658 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 2659 if( pExpr->iColumn<0 ) return 1; 2660 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2661 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 2662 } 2663 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 2664 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2665 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 2666 if( sqlite3ExprCompare(0, pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 2667 return 1; 2668 } 2669 } 2670 } 2671 } 2672 return 0; 2673 } 2674 2675 /* 2676 ** Return a bitmask where 1s indicate that the corresponding column of 2677 ** the table is used by an index. Only the first 63 columns are considered. 2678 */ 2679 static Bitmask columnsInIndex(Index *pIdx){ 2680 Bitmask m = 0; 2681 int j; 2682 for(j=pIdx->nColumn-1; j>=0; j--){ 2683 int x = pIdx->aiColumn[j]; 2684 if( x>=0 ){ 2685 testcase( x==BMS-1 ); 2686 testcase( x==BMS-2 ); 2687 if( x<BMS-1 ) m |= MASKBIT(x); 2688 } 2689 } 2690 return m; 2691 } 2692 2693 /* Check to see if a partial index with pPartIndexWhere can be used 2694 ** in the current query. Return true if it can be and false if not. 2695 */ 2696 static int whereUsablePartialIndex(int iTab, WhereClause *pWC, Expr *pWhere){ 2697 int i; 2698 WhereTerm *pTerm; 2699 Parse *pParse = pWC->pWInfo->pParse; 2700 while( pWhere->op==TK_AND ){ 2701 if( !whereUsablePartialIndex(iTab,pWC,pWhere->pLeft) ) return 0; 2702 pWhere = pWhere->pRight; 2703 } 2704 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 2705 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2706 Expr *pExpr = pTerm->pExpr; 2707 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 2708 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 2709 ){ 2710 return 1; 2711 } 2712 } 2713 return 0; 2714 } 2715 2716 /* 2717 ** Add all WhereLoop objects for a single table of the join where the table 2718 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 2719 ** a b-tree table, not a virtual table. 2720 ** 2721 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 2722 ** are calculated as follows: 2723 ** 2724 ** For a full scan, assuming the table (or index) contains nRow rows: 2725 ** 2726 ** cost = nRow * 3.0 // full-table scan 2727 ** cost = nRow * K // scan of covering index 2728 ** cost = nRow * (K+3.0) // scan of non-covering index 2729 ** 2730 ** where K is a value between 1.1 and 3.0 set based on the relative 2731 ** estimated average size of the index and table records. 2732 ** 2733 ** For an index scan, where nVisit is the number of index rows visited 2734 ** by the scan, and nSeek is the number of seek operations required on 2735 ** the index b-tree: 2736 ** 2737 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 2738 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 2739 ** 2740 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 2741 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 2742 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 2743 ** 2744 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 2745 ** of uncertainty. For this reason, scoring is designed to pick plans that 2746 ** "do the least harm" if the estimates are inaccurate. For example, a 2747 ** log(nRow) factor is omitted from a non-covering index scan in order to 2748 ** bias the scoring in favor of using an index, since the worst-case 2749 ** performance of using an index is far better than the worst-case performance 2750 ** of a full table scan. 2751 */ 2752 static int whereLoopAddBtree( 2753 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2754 Bitmask mPrereq /* Extra prerequesites for using this table */ 2755 ){ 2756 WhereInfo *pWInfo; /* WHERE analysis context */ 2757 Index *pProbe; /* An index we are evaluating */ 2758 Index sPk; /* A fake index object for the primary key */ 2759 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 2760 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 2761 SrcList *pTabList; /* The FROM clause */ 2762 struct SrcList_item *pSrc; /* The FROM clause btree term to add */ 2763 WhereLoop *pNew; /* Template WhereLoop object */ 2764 int rc = SQLITE_OK; /* Return code */ 2765 int iSortIdx = 1; /* Index number */ 2766 int b; /* A boolean value */ 2767 LogEst rSize; /* number of rows in the table */ 2768 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 2769 WhereClause *pWC; /* The parsed WHERE clause */ 2770 Table *pTab; /* Table being queried */ 2771 2772 pNew = pBuilder->pNew; 2773 pWInfo = pBuilder->pWInfo; 2774 pTabList = pWInfo->pTabList; 2775 pSrc = pTabList->a + pNew->iTab; 2776 pTab = pSrc->pTab; 2777 pWC = pBuilder->pWC; 2778 assert( !IsVirtual(pSrc->pTab) ); 2779 2780 if( pSrc->pIBIndex ){ 2781 /* An INDEXED BY clause specifies a particular index to use */ 2782 pProbe = pSrc->pIBIndex; 2783 }else if( !HasRowid(pTab) ){ 2784 pProbe = pTab->pIndex; 2785 }else{ 2786 /* There is no INDEXED BY clause. Create a fake Index object in local 2787 ** variable sPk to represent the rowid primary key index. Make this 2788 ** fake index the first in a chain of Index objects with all of the real 2789 ** indices to follow */ 2790 Index *pFirst; /* First of real indices on the table */ 2791 memset(&sPk, 0, sizeof(Index)); 2792 sPk.nKeyCol = 1; 2793 sPk.nColumn = 1; 2794 sPk.aiColumn = &aiColumnPk; 2795 sPk.aiRowLogEst = aiRowEstPk; 2796 sPk.onError = OE_Replace; 2797 sPk.pTable = pTab; 2798 sPk.szIdxRow = pTab->szTabRow; 2799 aiRowEstPk[0] = pTab->nRowLogEst; 2800 aiRowEstPk[1] = 0; 2801 pFirst = pSrc->pTab->pIndex; 2802 if( pSrc->fg.notIndexed==0 ){ 2803 /* The real indices of the table are only considered if the 2804 ** NOT INDEXED qualifier is omitted from the FROM clause */ 2805 sPk.pNext = pFirst; 2806 } 2807 pProbe = &sPk; 2808 } 2809 rSize = pTab->nRowLogEst; 2810 rLogSize = estLog(rSize); 2811 2812 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 2813 /* Automatic indexes */ 2814 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 2815 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 2816 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 2817 && pSrc->pIBIndex==0 /* Has no INDEXED BY clause */ 2818 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 2819 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 2820 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 2821 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 2822 ){ 2823 /* Generate auto-index WhereLoops */ 2824 WhereTerm *pTerm; 2825 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 2826 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 2827 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2828 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 2829 pNew->u.btree.nEq = 1; 2830 pNew->nSkip = 0; 2831 pNew->u.btree.pIndex = 0; 2832 pNew->nLTerm = 1; 2833 pNew->aLTerm[0] = pTerm; 2834 /* TUNING: One-time cost for computing the automatic index is 2835 ** estimated to be X*N*log2(N) where N is the number of rows in 2836 ** the table being indexed and where X is 7 (LogEst=28) for normal 2837 ** tables or 1.375 (LogEst=4) for views and subqueries. The value 2838 ** of X is smaller for views and subqueries so that the query planner 2839 ** will be more aggressive about generating automatic indexes for 2840 ** those objects, since there is no opportunity to add schema 2841 ** indexes on subqueries and views. */ 2842 pNew->rSetup = rLogSize + rSize + 4; 2843 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ 2844 pNew->rSetup += 24; 2845 } 2846 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 2847 if( pNew->rSetup<0 ) pNew->rSetup = 0; 2848 /* TUNING: Each index lookup yields 20 rows in the table. This 2849 ** is more than the usual guess of 10 rows, since we have no way 2850 ** of knowing how selective the index will ultimately be. It would 2851 ** not be unreasonable to make this value much larger. */ 2852 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 2853 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 2854 pNew->wsFlags = WHERE_AUTO_INDEX; 2855 pNew->prereq = mPrereq | pTerm->prereqRight; 2856 rc = whereLoopInsert(pBuilder, pNew); 2857 } 2858 } 2859 } 2860 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 2861 2862 /* Loop over all indices 2863 */ 2864 for(; rc==SQLITE_OK && pProbe; pProbe=pProbe->pNext, iSortIdx++){ 2865 if( pProbe->pPartIdxWhere!=0 2866 && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){ 2867 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 2868 continue; /* Partial index inappropriate for this query */ 2869 } 2870 rSize = pProbe->aiRowLogEst[0]; 2871 pNew->u.btree.nEq = 0; 2872 pNew->u.btree.nBtm = 0; 2873 pNew->u.btree.nTop = 0; 2874 pNew->nSkip = 0; 2875 pNew->nLTerm = 0; 2876 pNew->iSortIdx = 0; 2877 pNew->rSetup = 0; 2878 pNew->prereq = mPrereq; 2879 pNew->nOut = rSize; 2880 pNew->u.btree.pIndex = pProbe; 2881 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 2882 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 2883 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 2884 if( pProbe->tnum<=0 ){ 2885 /* Integer primary key index */ 2886 pNew->wsFlags = WHERE_IPK; 2887 2888 /* Full table scan */ 2889 pNew->iSortIdx = b ? iSortIdx : 0; 2890 /* TUNING: Cost of full table scan is (N*3.0). */ 2891 pNew->rRun = rSize + 16; 2892 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2893 whereLoopOutputAdjust(pWC, pNew, rSize); 2894 rc = whereLoopInsert(pBuilder, pNew); 2895 pNew->nOut = rSize; 2896 if( rc ) break; 2897 }else{ 2898 Bitmask m; 2899 if( pProbe->isCovering ){ 2900 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 2901 m = 0; 2902 }else{ 2903 m = pSrc->colUsed & ~columnsInIndex(pProbe); 2904 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 2905 } 2906 2907 /* Full scan via index */ 2908 if( b 2909 || !HasRowid(pTab) 2910 || pProbe->pPartIdxWhere!=0 2911 || ( m==0 2912 && pProbe->bUnordered==0 2913 && (pProbe->szIdxRow<pTab->szTabRow) 2914 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 2915 && sqlite3GlobalConfig.bUseCis 2916 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 2917 ) 2918 ){ 2919 pNew->iSortIdx = b ? iSortIdx : 0; 2920 2921 /* The cost of visiting the index rows is N*K, where K is 2922 ** between 1.1 and 3.0, depending on the relative sizes of the 2923 ** index and table rows. */ 2924 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 2925 if( m!=0 ){ 2926 /* If this is a non-covering index scan, add in the cost of 2927 ** doing table lookups. The cost will be 3x the number of 2928 ** lookups. Take into account WHERE clause terms that can be 2929 ** satisfied using just the index, and that do not require a 2930 ** table lookup. */ 2931 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 2932 int ii; 2933 int iCur = pSrc->iCursor; 2934 WhereClause *pWC2 = &pWInfo->sWC; 2935 for(ii=0; ii<pWC2->nTerm; ii++){ 2936 WhereTerm *pTerm = &pWC2->a[ii]; 2937 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 2938 break; 2939 } 2940 /* pTerm can be evaluated using just the index. So reduce 2941 ** the expected number of table lookups accordingly */ 2942 if( pTerm->truthProb<=0 ){ 2943 nLookup += pTerm->truthProb; 2944 }else{ 2945 nLookup--; 2946 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 2947 } 2948 } 2949 2950 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 2951 } 2952 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2953 whereLoopOutputAdjust(pWC, pNew, rSize); 2954 rc = whereLoopInsert(pBuilder, pNew); 2955 pNew->nOut = rSize; 2956 if( rc ) break; 2957 } 2958 } 2959 2960 pBuilder->bldFlags = 0; 2961 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 2962 if( pBuilder->bldFlags==SQLITE_BLDF_INDEXED ){ 2963 /* If a non-unique index is used, or if a prefix of the key for 2964 ** unique index is used (making the index functionally non-unique) 2965 ** then the sqlite_stat1 data becomes important for scoring the 2966 ** plan */ 2967 pTab->tabFlags |= TF_StatsUsed; 2968 } 2969 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2970 sqlite3Stat4ProbeFree(pBuilder->pRec); 2971 pBuilder->nRecValid = 0; 2972 pBuilder->pRec = 0; 2973 #endif 2974 2975 /* If there was an INDEXED BY clause, then only that one index is 2976 ** considered. */ 2977 if( pSrc->pIBIndex ) break; 2978 } 2979 return rc; 2980 } 2981 2982 #ifndef SQLITE_OMIT_VIRTUALTABLE 2983 2984 /* 2985 ** Argument pIdxInfo is already populated with all constraints that may 2986 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 2987 ** function marks a subset of those constraints usable, invokes the 2988 ** xBestIndex method and adds the returned plan to pBuilder. 2989 ** 2990 ** A constraint is marked usable if: 2991 ** 2992 ** * Argument mUsable indicates that its prerequisites are available, and 2993 ** 2994 ** * It is not one of the operators specified in the mExclude mask passed 2995 ** as the fourth argument (which in practice is either WO_IN or 0). 2996 ** 2997 ** Argument mPrereq is a mask of tables that must be scanned before the 2998 ** virtual table in question. These are added to the plans prerequisites 2999 ** before it is added to pBuilder. 3000 ** 3001 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3002 ** uses one or more WO_IN terms, or false otherwise. 3003 */ 3004 static int whereLoopAddVirtualOne( 3005 WhereLoopBuilder *pBuilder, 3006 Bitmask mPrereq, /* Mask of tables that must be used. */ 3007 Bitmask mUsable, /* Mask of usable tables */ 3008 u16 mExclude, /* Exclude terms using these operators */ 3009 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3010 u16 mNoOmit, /* Do not omit these constraints */ 3011 int *pbIn /* OUT: True if plan uses an IN(...) op */ 3012 ){ 3013 WhereClause *pWC = pBuilder->pWC; 3014 struct sqlite3_index_constraint *pIdxCons; 3015 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3016 int i; 3017 int mxTerm; 3018 int rc = SQLITE_OK; 3019 WhereLoop *pNew = pBuilder->pNew; 3020 Parse *pParse = pBuilder->pWInfo->pParse; 3021 struct SrcList_item *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3022 int nConstraint = pIdxInfo->nConstraint; 3023 3024 assert( (mUsable & mPrereq)==mPrereq ); 3025 *pbIn = 0; 3026 pNew->prereq = mPrereq; 3027 3028 /* Set the usable flag on the subset of constraints identified by 3029 ** arguments mUsable and mExclude. */ 3030 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3031 for(i=0; i<nConstraint; i++, pIdxCons++){ 3032 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3033 pIdxCons->usable = 0; 3034 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3035 && (pTerm->eOperator & mExclude)==0 3036 ){ 3037 pIdxCons->usable = 1; 3038 } 3039 } 3040 3041 /* Initialize the output fields of the sqlite3_index_info structure */ 3042 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3043 assert( pIdxInfo->needToFreeIdxStr==0 ); 3044 pIdxInfo->idxStr = 0; 3045 pIdxInfo->idxNum = 0; 3046 pIdxInfo->orderByConsumed = 0; 3047 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3048 pIdxInfo->estimatedRows = 25; 3049 pIdxInfo->idxFlags = 0; 3050 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3051 3052 /* Invoke the virtual table xBestIndex() method */ 3053 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3054 if( rc ) return rc; 3055 3056 mxTerm = -1; 3057 assert( pNew->nLSlot>=nConstraint ); 3058 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 3059 pNew->u.vtab.omitMask = 0; 3060 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3061 for(i=0; i<nConstraint; i++, pIdxCons++){ 3062 int iTerm; 3063 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3064 WhereTerm *pTerm; 3065 int j = pIdxCons->iTermOffset; 3066 if( iTerm>=nConstraint 3067 || j<0 3068 || j>=pWC->nTerm 3069 || pNew->aLTerm[iTerm]!=0 3070 || pIdxCons->usable==0 3071 ){ 3072 rc = SQLITE_ERROR; 3073 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3074 return rc; 3075 } 3076 testcase( iTerm==nConstraint-1 ); 3077 testcase( j==0 ); 3078 testcase( j==pWC->nTerm-1 ); 3079 pTerm = &pWC->a[j]; 3080 pNew->prereq |= pTerm->prereqRight; 3081 assert( iTerm<pNew->nLSlot ); 3082 pNew->aLTerm[iTerm] = pTerm; 3083 if( iTerm>mxTerm ) mxTerm = iTerm; 3084 testcase( iTerm==15 ); 3085 testcase( iTerm==16 ); 3086 if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<<iTerm; 3087 if( (pTerm->eOperator & WO_IN)!=0 ){ 3088 /* A virtual table that is constrained by an IN clause may not 3089 ** consume the ORDER BY clause because (1) the order of IN terms 3090 ** is not necessarily related to the order of output terms and 3091 ** (2) Multiple outputs from a single IN value will not merge 3092 ** together. */ 3093 pIdxInfo->orderByConsumed = 0; 3094 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3095 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3096 } 3097 } 3098 } 3099 pNew->u.vtab.omitMask &= ~mNoOmit; 3100 3101 pNew->nLTerm = mxTerm+1; 3102 assert( pNew->nLTerm<=pNew->nLSlot ); 3103 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3104 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3105 pIdxInfo->needToFreeIdxStr = 0; 3106 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3107 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3108 pIdxInfo->nOrderBy : 0); 3109 pNew->rSetup = 0; 3110 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3111 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3112 3113 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3114 ** that the scan will visit at most one row. Clear it otherwise. */ 3115 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3116 pNew->wsFlags |= WHERE_ONEROW; 3117 }else{ 3118 pNew->wsFlags &= ~WHERE_ONEROW; 3119 } 3120 rc = whereLoopInsert(pBuilder, pNew); 3121 if( pNew->u.vtab.needFree ){ 3122 sqlite3_free(pNew->u.vtab.idxStr); 3123 pNew->u.vtab.needFree = 0; 3124 } 3125 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3126 *pbIn, (sqlite3_uint64)mPrereq, 3127 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3128 3129 return rc; 3130 } 3131 3132 3133 /* 3134 ** Add all WhereLoop objects for a table of the join identified by 3135 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3136 ** 3137 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3138 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3139 ** entries that occur before the virtual table in the FROM clause and are 3140 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3141 ** mUnusable mask contains all FROM clause entries that occur after the 3142 ** virtual table and are separated from it by at least one LEFT or 3143 ** CROSS JOIN. 3144 ** 3145 ** For example, if the query were: 3146 ** 3147 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3148 ** 3149 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3150 ** 3151 ** All the tables in mPrereq must be scanned before the current virtual 3152 ** table. So any terms for which all prerequisites are satisfied by 3153 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3154 ** Conversely, all tables in mUnusable must be scanned after the current 3155 ** virtual table, so any terms for which the prerequisites overlap with 3156 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3157 */ 3158 static int whereLoopAddVirtual( 3159 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3160 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3161 Bitmask mUnusable /* Tables that must be scanned after this one */ 3162 ){ 3163 int rc = SQLITE_OK; /* Return code */ 3164 WhereInfo *pWInfo; /* WHERE analysis context */ 3165 Parse *pParse; /* The parsing context */ 3166 WhereClause *pWC; /* The WHERE clause */ 3167 struct SrcList_item *pSrc; /* The FROM clause term to search */ 3168 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3169 int nConstraint; /* Number of constraints in p */ 3170 int bIn; /* True if plan uses IN(...) operator */ 3171 WhereLoop *pNew; 3172 Bitmask mBest; /* Tables used by best possible plan */ 3173 u16 mNoOmit; 3174 3175 assert( (mPrereq & mUnusable)==0 ); 3176 pWInfo = pBuilder->pWInfo; 3177 pParse = pWInfo->pParse; 3178 pWC = pBuilder->pWC; 3179 pNew = pBuilder->pNew; 3180 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3181 assert( IsVirtual(pSrc->pTab) ); 3182 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy, 3183 &mNoOmit); 3184 if( p==0 ) return SQLITE_NOMEM_BKPT; 3185 pNew->rSetup = 0; 3186 pNew->wsFlags = WHERE_VIRTUALTABLE; 3187 pNew->nLTerm = 0; 3188 pNew->u.vtab.needFree = 0; 3189 nConstraint = p->nConstraint; 3190 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3191 sqlite3DbFree(pParse->db, p); 3192 return SQLITE_NOMEM_BKPT; 3193 } 3194 3195 /* First call xBestIndex() with all constraints usable. */ 3196 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3197 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn); 3198 3199 /* If the call to xBestIndex() with all terms enabled produced a plan 3200 ** that does not require any source tables (IOW: a plan with mBest==0), 3201 ** then there is no point in making any further calls to xBestIndex() 3202 ** since they will all return the same result (if the xBestIndex() 3203 ** implementation is sane). */ 3204 if( rc==SQLITE_OK && (mBest = (pNew->prereq & ~mPrereq))!=0 ){ 3205 int seenZero = 0; /* True if a plan with no prereqs seen */ 3206 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3207 Bitmask mPrev = 0; 3208 Bitmask mBestNoIn = 0; 3209 3210 /* If the plan produced by the earlier call uses an IN(...) term, call 3211 ** xBestIndex again, this time with IN(...) terms disabled. */ 3212 if( bIn ){ 3213 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3214 rc = whereLoopAddVirtualOne( 3215 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn); 3216 assert( bIn==0 ); 3217 mBestNoIn = pNew->prereq & ~mPrereq; 3218 if( mBestNoIn==0 ){ 3219 seenZero = 1; 3220 seenZeroNoIN = 1; 3221 } 3222 } 3223 3224 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3225 ** in the set of terms that apply to the current virtual table. */ 3226 while( rc==SQLITE_OK ){ 3227 int i; 3228 Bitmask mNext = ALLBITS; 3229 assert( mNext>0 ); 3230 for(i=0; i<nConstraint; i++){ 3231 Bitmask mThis = ( 3232 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3233 ); 3234 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3235 } 3236 mPrev = mNext; 3237 if( mNext==ALLBITS ) break; 3238 if( mNext==mBest || mNext==mBestNoIn ) continue; 3239 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3240 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3241 rc = whereLoopAddVirtualOne( 3242 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn); 3243 if( pNew->prereq==mPrereq ){ 3244 seenZero = 1; 3245 if( bIn==0 ) seenZeroNoIN = 1; 3246 } 3247 } 3248 3249 /* If the calls to xBestIndex() in the above loop did not find a plan 3250 ** that requires no source tables at all (i.e. one guaranteed to be 3251 ** usable), make a call here with all source tables disabled */ 3252 if( rc==SQLITE_OK && seenZero==0 ){ 3253 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3254 rc = whereLoopAddVirtualOne( 3255 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn); 3256 if( bIn==0 ) seenZeroNoIN = 1; 3257 } 3258 3259 /* If the calls to xBestIndex() have so far failed to find a plan 3260 ** that requires no source tables at all and does not use an IN(...) 3261 ** operator, make a final call to obtain one here. */ 3262 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3263 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3264 rc = whereLoopAddVirtualOne( 3265 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn); 3266 } 3267 } 3268 3269 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3270 sqlite3DbFreeNN(pParse->db, p); 3271 return rc; 3272 } 3273 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3274 3275 /* 3276 ** Add WhereLoop entries to handle OR terms. This works for either 3277 ** btrees or virtual tables. 3278 */ 3279 static int whereLoopAddOr( 3280 WhereLoopBuilder *pBuilder, 3281 Bitmask mPrereq, 3282 Bitmask mUnusable 3283 ){ 3284 WhereInfo *pWInfo = pBuilder->pWInfo; 3285 WhereClause *pWC; 3286 WhereLoop *pNew; 3287 WhereTerm *pTerm, *pWCEnd; 3288 int rc = SQLITE_OK; 3289 int iCur; 3290 WhereClause tempWC; 3291 WhereLoopBuilder sSubBuild; 3292 WhereOrSet sSum, sCur; 3293 struct SrcList_item *pItem; 3294 3295 pWC = pBuilder->pWC; 3296 pWCEnd = pWC->a + pWC->nTerm; 3297 pNew = pBuilder->pNew; 3298 memset(&sSum, 0, sizeof(sSum)); 3299 pItem = pWInfo->pTabList->a + pNew->iTab; 3300 iCur = pItem->iCursor; 3301 3302 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3303 if( (pTerm->eOperator & WO_OR)!=0 3304 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3305 ){ 3306 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3307 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3308 WhereTerm *pOrTerm; 3309 int once = 1; 3310 int i, j; 3311 3312 sSubBuild = *pBuilder; 3313 sSubBuild.pOrderBy = 0; 3314 sSubBuild.pOrSet = &sCur; 3315 3316 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3317 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3318 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3319 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3320 }else if( pOrTerm->leftCursor==iCur ){ 3321 tempWC.pWInfo = pWC->pWInfo; 3322 tempWC.pOuter = pWC; 3323 tempWC.op = TK_AND; 3324 tempWC.nTerm = 1; 3325 tempWC.a = pOrTerm; 3326 sSubBuild.pWC = &tempWC; 3327 }else{ 3328 continue; 3329 } 3330 sCur.n = 0; 3331 #ifdef WHERETRACE_ENABLED 3332 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3333 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3334 if( sqlite3WhereTrace & 0x400 ){ 3335 sqlite3WhereClausePrint(sSubBuild.pWC); 3336 } 3337 #endif 3338 #ifndef SQLITE_OMIT_VIRTUALTABLE 3339 if( IsVirtual(pItem->pTab) ){ 3340 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 3341 }else 3342 #endif 3343 { 3344 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 3345 } 3346 if( rc==SQLITE_OK ){ 3347 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 3348 } 3349 assert( rc==SQLITE_OK || sCur.n==0 ); 3350 if( sCur.n==0 ){ 3351 sSum.n = 0; 3352 break; 3353 }else if( once ){ 3354 whereOrMove(&sSum, &sCur); 3355 once = 0; 3356 }else{ 3357 WhereOrSet sPrev; 3358 whereOrMove(&sPrev, &sSum); 3359 sSum.n = 0; 3360 for(i=0; i<sPrev.n; i++){ 3361 for(j=0; j<sCur.n; j++){ 3362 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3363 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3364 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3365 } 3366 } 3367 } 3368 } 3369 pNew->nLTerm = 1; 3370 pNew->aLTerm[0] = pTerm; 3371 pNew->wsFlags = WHERE_MULTI_OR; 3372 pNew->rSetup = 0; 3373 pNew->iSortIdx = 0; 3374 memset(&pNew->u, 0, sizeof(pNew->u)); 3375 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3376 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3377 ** of all sub-scans required by the OR-scan. However, due to rounding 3378 ** errors, it may be that the cost of the OR-scan is equal to its 3379 ** most expensive sub-scan. Add the smallest possible penalty 3380 ** (equivalent to multiplying the cost by 1.07) to ensure that 3381 ** this does not happen. Otherwise, for WHERE clauses such as the 3382 ** following where there is an index on "y": 3383 ** 3384 ** WHERE likelihood(x=?, 0.99) OR y=? 3385 ** 3386 ** the planner may elect to "OR" together a full-table scan and an 3387 ** index lookup. And other similarly odd results. */ 3388 pNew->rRun = sSum.a[i].rRun + 1; 3389 pNew->nOut = sSum.a[i].nOut; 3390 pNew->prereq = sSum.a[i].prereq; 3391 rc = whereLoopInsert(pBuilder, pNew); 3392 } 3393 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3394 } 3395 } 3396 return rc; 3397 } 3398 3399 /* 3400 ** Add all WhereLoop objects for all tables 3401 */ 3402 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3403 WhereInfo *pWInfo = pBuilder->pWInfo; 3404 Bitmask mPrereq = 0; 3405 Bitmask mPrior = 0; 3406 int iTab; 3407 SrcList *pTabList = pWInfo->pTabList; 3408 struct SrcList_item *pItem; 3409 struct SrcList_item *pEnd = &pTabList->a[pWInfo->nLevel]; 3410 sqlite3 *db = pWInfo->pParse->db; 3411 int rc = SQLITE_OK; 3412 WhereLoop *pNew; 3413 u8 priorJointype = 0; 3414 3415 /* Loop over the tables in the join, from left to right */ 3416 pNew = pBuilder->pNew; 3417 whereLoopInit(pNew); 3418 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3419 Bitmask mUnusable = 0; 3420 pNew->iTab = iTab; 3421 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3422 if( ((pItem->fg.jointype|priorJointype) & (JT_LEFT|JT_CROSS))!=0 ){ 3423 /* This condition is true when pItem is the FROM clause term on the 3424 ** right-hand-side of a LEFT or CROSS JOIN. */ 3425 mPrereq = mPrior; 3426 } 3427 priorJointype = pItem->fg.jointype; 3428 #ifndef SQLITE_OMIT_VIRTUALTABLE 3429 if( IsVirtual(pItem->pTab) ){ 3430 struct SrcList_item *p; 3431 for(p=&pItem[1]; p<pEnd; p++){ 3432 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3433 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3434 } 3435 } 3436 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 3437 }else 3438 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3439 { 3440 rc = whereLoopAddBtree(pBuilder, mPrereq); 3441 } 3442 if( rc==SQLITE_OK ){ 3443 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 3444 } 3445 mPrior |= pNew->maskSelf; 3446 if( rc || db->mallocFailed ) break; 3447 } 3448 3449 whereLoopClear(db, pNew); 3450 return rc; 3451 } 3452 3453 /* 3454 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 3455 ** parameters) to see if it outputs rows in the requested ORDER BY 3456 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3457 ** 3458 ** N>0: N terms of the ORDER BY clause are satisfied 3459 ** N==0: No terms of the ORDER BY clause are satisfied 3460 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3461 ** 3462 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3463 ** strict. With GROUP BY and DISTINCT the only requirement is that 3464 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3465 ** and DISTINCT do not require rows to appear in any particular order as long 3466 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3467 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3468 ** pOrderBy terms must be matched in strict left-to-right order. 3469 */ 3470 static i8 wherePathSatisfiesOrderBy( 3471 WhereInfo *pWInfo, /* The WHERE clause */ 3472 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3473 WherePath *pPath, /* The WherePath to check */ 3474 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 3475 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3476 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3477 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3478 ){ 3479 u8 revSet; /* True if rev is known */ 3480 u8 rev; /* Composite sort order */ 3481 u8 revIdx; /* Index sort order */ 3482 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 3483 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 3484 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 3485 u16 eqOpMask; /* Allowed equality operators */ 3486 u16 nKeyCol; /* Number of key columns in pIndex */ 3487 u16 nColumn; /* Total number of ordered columns in the index */ 3488 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 3489 int iLoop; /* Index of WhereLoop in pPath being processed */ 3490 int i, j; /* Loop counters */ 3491 int iCur; /* Cursor number for current WhereLoop */ 3492 int iColumn; /* A column number within table iCur */ 3493 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 3494 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3495 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 3496 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 3497 Index *pIndex; /* The index associated with pLoop */ 3498 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 3499 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 3500 Bitmask obDone; /* Mask of all ORDER BY terms */ 3501 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 3502 Bitmask ready; /* Mask of inner loops */ 3503 3504 /* 3505 ** We say the WhereLoop is "one-row" if it generates no more than one 3506 ** row of output. A WhereLoop is one-row if all of the following are true: 3507 ** (a) All index columns match with WHERE_COLUMN_EQ. 3508 ** (b) The index is unique 3509 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 3510 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 3511 ** 3512 ** We say the WhereLoop is "order-distinct" if the set of columns from 3513 ** that WhereLoop that are in the ORDER BY clause are different for every 3514 ** row of the WhereLoop. Every one-row WhereLoop is automatically 3515 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 3516 ** is not order-distinct. To be order-distinct is not quite the same as being 3517 ** UNIQUE since a UNIQUE column or index can have multiple rows that 3518 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 3519 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 3520 ** 3521 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 3522 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 3523 ** automatically order-distinct. 3524 */ 3525 3526 assert( pOrderBy!=0 ); 3527 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 3528 3529 nOrderBy = pOrderBy->nExpr; 3530 testcase( nOrderBy==BMS-1 ); 3531 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 3532 isOrderDistinct = 1; 3533 obDone = MASKBIT(nOrderBy)-1; 3534 orderDistinctMask = 0; 3535 ready = 0; 3536 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 3537 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) eqOpMask |= WO_IN; 3538 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 3539 if( iLoop>0 ) ready |= pLoop->maskSelf; 3540 if( iLoop<nLoop ){ 3541 pLoop = pPath->aLoop[iLoop]; 3542 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 3543 }else{ 3544 pLoop = pLast; 3545 } 3546 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 3547 if( pLoop->u.vtab.isOrdered ) obSat = obDone; 3548 break; 3549 }else{ 3550 pLoop->u.btree.nIdxCol = 0; 3551 } 3552 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 3553 3554 /* Mark off any ORDER BY term X that is a column in the table of 3555 ** the current loop for which there is term in the WHERE 3556 ** clause of the form X IS NULL or X=? that reference only outer 3557 ** loops. 3558 */ 3559 for(i=0; i<nOrderBy; i++){ 3560 if( MASKBIT(i) & obSat ) continue; 3561 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3562 if( pOBExpr->op!=TK_COLUMN ) continue; 3563 if( pOBExpr->iTable!=iCur ) continue; 3564 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 3565 ~ready, eqOpMask, 0); 3566 if( pTerm==0 ) continue; 3567 if( pTerm->eOperator==WO_IN ){ 3568 /* IN terms are only valid for sorting in the ORDER BY LIMIT 3569 ** optimization, and then only if they are actually used 3570 ** by the query plan */ 3571 assert( wctrlFlags & WHERE_ORDERBY_LIMIT ); 3572 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 3573 if( j>=pLoop->nLTerm ) continue; 3574 } 3575 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 3576 const char *z1, *z2; 3577 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3578 if( !pColl ) pColl = db->pDfltColl; 3579 z1 = pColl->zName; 3580 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pTerm->pExpr); 3581 if( !pColl ) pColl = db->pDfltColl; 3582 z2 = pColl->zName; 3583 if( sqlite3StrICmp(z1, z2)!=0 ) continue; 3584 testcase( pTerm->pExpr->op==TK_IS ); 3585 } 3586 obSat |= MASKBIT(i); 3587 } 3588 3589 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 3590 if( pLoop->wsFlags & WHERE_IPK ){ 3591 pIndex = 0; 3592 nKeyCol = 0; 3593 nColumn = 1; 3594 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 3595 return 0; 3596 }else{ 3597 nKeyCol = pIndex->nKeyCol; 3598 nColumn = pIndex->nColumn; 3599 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 3600 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 3601 || !HasRowid(pIndex->pTable)); 3602 isOrderDistinct = IsUniqueIndex(pIndex); 3603 } 3604 3605 /* Loop through all columns of the index and deal with the ones 3606 ** that are not constrained by == or IN. 3607 */ 3608 rev = revSet = 0; 3609 distinctColumns = 0; 3610 for(j=0; j<nColumn; j++){ 3611 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 3612 3613 assert( j>=pLoop->u.btree.nEq 3614 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 3615 ); 3616 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 3617 u16 eOp = pLoop->aLTerm[j]->eOperator; 3618 3619 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 3620 ** doing WHERE_ORDERBY_LIMIT processing). 3621 ** 3622 ** If the current term is a column of an ((?,?) IN (SELECT...)) 3623 ** expression for which the SELECT returns more than one column, 3624 ** check that it is the only column used by this loop. Otherwise, 3625 ** if it is one of two or more, none of the columns can be 3626 ** considered to match an ORDER BY term. */ 3627 if( (eOp & eqOpMask)!=0 ){ 3628 if( eOp & WO_ISNULL ){ 3629 testcase( isOrderDistinct ); 3630 isOrderDistinct = 0; 3631 } 3632 continue; 3633 }else if( ALWAYS(eOp & WO_IN) ){ 3634 /* ALWAYS() justification: eOp is an equality operator due to the 3635 ** j<pLoop->u.btree.nEq constraint above. Any equality other 3636 ** than WO_IN is captured by the previous "if". So this one 3637 ** always has to be WO_IN. */ 3638 Expr *pX = pLoop->aLTerm[j]->pExpr; 3639 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 3640 if( pLoop->aLTerm[i]->pExpr==pX ){ 3641 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 3642 bOnce = 0; 3643 break; 3644 } 3645 } 3646 } 3647 } 3648 3649 /* Get the column number in the table (iColumn) and sort order 3650 ** (revIdx) for the j-th column of the index. 3651 */ 3652 if( pIndex ){ 3653 iColumn = pIndex->aiColumn[j]; 3654 revIdx = pIndex->aSortOrder[j]; 3655 if( iColumn==pIndex->pTable->iPKey ) iColumn = -1; 3656 }else{ 3657 iColumn = XN_ROWID; 3658 revIdx = 0; 3659 } 3660 3661 /* An unconstrained column that might be NULL means that this 3662 ** WhereLoop is not well-ordered 3663 */ 3664 if( isOrderDistinct 3665 && iColumn>=0 3666 && j>=pLoop->u.btree.nEq 3667 && pIndex->pTable->aCol[iColumn].notNull==0 3668 ){ 3669 isOrderDistinct = 0; 3670 } 3671 3672 /* Find the ORDER BY term that corresponds to the j-th column 3673 ** of the index and mark that ORDER BY term off 3674 */ 3675 isMatch = 0; 3676 for(i=0; bOnce && i<nOrderBy; i++){ 3677 if( MASKBIT(i) & obSat ) continue; 3678 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3679 testcase( wctrlFlags & WHERE_GROUPBY ); 3680 testcase( wctrlFlags & WHERE_DISTINCTBY ); 3681 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 3682 if( iColumn>=(-1) ){ 3683 if( pOBExpr->op!=TK_COLUMN ) continue; 3684 if( pOBExpr->iTable!=iCur ) continue; 3685 if( pOBExpr->iColumn!=iColumn ) continue; 3686 }else{ 3687 if( sqlite3ExprCompare(0, 3688 pOBExpr,pIndex->aColExpr->a[j].pExpr,iCur) ){ 3689 continue; 3690 } 3691 } 3692 if( iColumn>=0 ){ 3693 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3694 if( !pColl ) pColl = db->pDfltColl; 3695 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 3696 } 3697 pLoop->u.btree.nIdxCol = j+1; 3698 isMatch = 1; 3699 break; 3700 } 3701 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 3702 /* Make sure the sort order is compatible in an ORDER BY clause. 3703 ** Sort order is irrelevant for a GROUP BY clause. */ 3704 if( revSet ){ 3705 if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0; 3706 }else{ 3707 rev = revIdx ^ pOrderBy->a[i].sortOrder; 3708 if( rev ) *pRevMask |= MASKBIT(iLoop); 3709 revSet = 1; 3710 } 3711 } 3712 if( isMatch ){ 3713 if( iColumn==XN_ROWID ){ 3714 testcase( distinctColumns==0 ); 3715 distinctColumns = 1; 3716 } 3717 obSat |= MASKBIT(i); 3718 }else{ 3719 /* No match found */ 3720 if( j==0 || j<nKeyCol ){ 3721 testcase( isOrderDistinct!=0 ); 3722 isOrderDistinct = 0; 3723 } 3724 break; 3725 } 3726 } /* end Loop over all index columns */ 3727 if( distinctColumns ){ 3728 testcase( isOrderDistinct==0 ); 3729 isOrderDistinct = 1; 3730 } 3731 } /* end-if not one-row */ 3732 3733 /* Mark off any other ORDER BY terms that reference pLoop */ 3734 if( isOrderDistinct ){ 3735 orderDistinctMask |= pLoop->maskSelf; 3736 for(i=0; i<nOrderBy; i++){ 3737 Expr *p; 3738 Bitmask mTerm; 3739 if( MASKBIT(i) & obSat ) continue; 3740 p = pOrderBy->a[i].pExpr; 3741 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 3742 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 3743 if( (mTerm&~orderDistinctMask)==0 ){ 3744 obSat |= MASKBIT(i); 3745 } 3746 } 3747 } 3748 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 3749 if( obSat==obDone ) return (i8)nOrderBy; 3750 if( !isOrderDistinct ){ 3751 for(i=nOrderBy-1; i>0; i--){ 3752 Bitmask m = MASKBIT(i) - 1; 3753 if( (obSat&m)==m ) return i; 3754 } 3755 return 0; 3756 } 3757 return -1; 3758 } 3759 3760 3761 /* 3762 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 3763 ** the planner assumes that the specified pOrderBy list is actually a GROUP 3764 ** BY clause - and so any order that groups rows as required satisfies the 3765 ** request. 3766 ** 3767 ** Normally, in this case it is not possible for the caller to determine 3768 ** whether or not the rows are really being delivered in sorted order, or 3769 ** just in some other order that provides the required grouping. However, 3770 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 3771 ** this function may be called on the returned WhereInfo object. It returns 3772 ** true if the rows really will be sorted in the specified order, or false 3773 ** otherwise. 3774 ** 3775 ** For example, assuming: 3776 ** 3777 ** CREATE INDEX i1 ON t1(x, Y); 3778 ** 3779 ** then 3780 ** 3781 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 3782 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 3783 */ 3784 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 3785 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 3786 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 3787 return pWInfo->sorted; 3788 } 3789 3790 #ifdef WHERETRACE_ENABLED 3791 /* For debugging use only: */ 3792 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 3793 static char zName[65]; 3794 int i; 3795 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 3796 if( pLast ) zName[i++] = pLast->cId; 3797 zName[i] = 0; 3798 return zName; 3799 } 3800 #endif 3801 3802 /* 3803 ** Return the cost of sorting nRow rows, assuming that the keys have 3804 ** nOrderby columns and that the first nSorted columns are already in 3805 ** order. 3806 */ 3807 static LogEst whereSortingCost( 3808 WhereInfo *pWInfo, 3809 LogEst nRow, 3810 int nOrderBy, 3811 int nSorted 3812 ){ 3813 /* TUNING: Estimated cost of a full external sort, where N is 3814 ** the number of rows to sort is: 3815 ** 3816 ** cost = (3.0 * N * log(N)). 3817 ** 3818 ** Or, if the order-by clause has X terms but only the last Y 3819 ** terms are out of order, then block-sorting will reduce the 3820 ** sorting cost to: 3821 ** 3822 ** cost = (3.0 * N * log(N)) * (Y/X) 3823 ** 3824 ** The (Y/X) term is implemented using stack variable rScale 3825 ** below. */ 3826 LogEst rScale, rSortCost; 3827 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 3828 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 3829 rSortCost = nRow + rScale + 16; 3830 3831 /* Multiple by log(M) where M is the number of output rows. 3832 ** Use the LIMIT for M if it is smaller */ 3833 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 3834 nRow = pWInfo->iLimit; 3835 } 3836 rSortCost += estLog(nRow); 3837 return rSortCost; 3838 } 3839 3840 /* 3841 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 3842 ** attempts to find the lowest cost path that visits each WhereLoop 3843 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 3844 ** 3845 ** Assume that the total number of output rows that will need to be sorted 3846 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 3847 ** costs if nRowEst==0. 3848 ** 3849 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 3850 ** error occurs. 3851 */ 3852 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 3853 int mxChoice; /* Maximum number of simultaneous paths tracked */ 3854 int nLoop; /* Number of terms in the join */ 3855 Parse *pParse; /* Parsing context */ 3856 sqlite3 *db; /* The database connection */ 3857 int iLoop; /* Loop counter over the terms of the join */ 3858 int ii, jj; /* Loop counters */ 3859 int mxI = 0; /* Index of next entry to replace */ 3860 int nOrderBy; /* Number of ORDER BY clause terms */ 3861 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 3862 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 3863 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 3864 WherePath *aFrom; /* All nFrom paths at the previous level */ 3865 WherePath *aTo; /* The nTo best paths at the current level */ 3866 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 3867 WherePath *pTo; /* An element of aTo[] that we are working on */ 3868 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 3869 WhereLoop **pX; /* Used to divy up the pSpace memory */ 3870 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 3871 char *pSpace; /* Temporary memory used by this routine */ 3872 int nSpace; /* Bytes of space allocated at pSpace */ 3873 3874 pParse = pWInfo->pParse; 3875 db = pParse->db; 3876 nLoop = pWInfo->nLevel; 3877 /* TUNING: For simple queries, only the best path is tracked. 3878 ** For 2-way joins, the 5 best paths are followed. 3879 ** For joins of 3 or more tables, track the 10 best paths */ 3880 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 3881 assert( nLoop<=pWInfo->pTabList->nSrc ); 3882 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 3883 3884 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 3885 ** case the purpose of this call is to estimate the number of rows returned 3886 ** by the overall query. Once this estimate has been obtained, the caller 3887 ** will invoke this function a second time, passing the estimate as the 3888 ** nRowEst parameter. */ 3889 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 3890 nOrderBy = 0; 3891 }else{ 3892 nOrderBy = pWInfo->pOrderBy->nExpr; 3893 } 3894 3895 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 3896 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 3897 nSpace += sizeof(LogEst) * nOrderBy; 3898 pSpace = sqlite3DbMallocRawNN(db, nSpace); 3899 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 3900 aTo = (WherePath*)pSpace; 3901 aFrom = aTo+mxChoice; 3902 memset(aFrom, 0, sizeof(aFrom[0])); 3903 pX = (WhereLoop**)(aFrom+mxChoice); 3904 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 3905 pFrom->aLoop = pX; 3906 } 3907 if( nOrderBy ){ 3908 /* If there is an ORDER BY clause and it is not being ignored, set up 3909 ** space for the aSortCost[] array. Each element of the aSortCost array 3910 ** is either zero - meaning it has not yet been initialized - or the 3911 ** cost of sorting nRowEst rows of data where the first X terms of 3912 ** the ORDER BY clause are already in order, where X is the array 3913 ** index. */ 3914 aSortCost = (LogEst*)pX; 3915 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 3916 } 3917 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 3918 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 3919 3920 /* Seed the search with a single WherePath containing zero WhereLoops. 3921 ** 3922 ** TUNING: Do not let the number of iterations go above 28. If the cost 3923 ** of computing an automatic index is not paid back within the first 28 3924 ** rows, then do not use the automatic index. */ 3925 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 3926 nFrom = 1; 3927 assert( aFrom[0].isOrdered==0 ); 3928 if( nOrderBy ){ 3929 /* If nLoop is zero, then there are no FROM terms in the query. Since 3930 ** in this case the query may return a maximum of one row, the results 3931 ** are already in the requested order. Set isOrdered to nOrderBy to 3932 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 3933 ** -1, indicating that the result set may or may not be ordered, 3934 ** depending on the loops added to the current plan. */ 3935 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 3936 } 3937 3938 /* Compute successively longer WherePaths using the previous generation 3939 ** of WherePaths as the basis for the next. Keep track of the mxChoice 3940 ** best paths at each generation */ 3941 for(iLoop=0; iLoop<nLoop; iLoop++){ 3942 nTo = 0; 3943 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 3944 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 3945 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 3946 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 3947 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 3948 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 3949 Bitmask maskNew; /* Mask of src visited by (..) */ 3950 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 3951 3952 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 3953 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 3954 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<10 ){ 3955 /* Do not use an automatic index if the this loop is expected 3956 ** to run less than 2 times. */ 3957 assert( 10==sqlite3LogEst(2) ); 3958 continue; 3959 } 3960 /* At this point, pWLoop is a candidate to be the next loop. 3961 ** Compute its cost */ 3962 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 3963 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 3964 nOut = pFrom->nRow + pWLoop->nOut; 3965 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 3966 if( isOrdered<0 ){ 3967 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 3968 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 3969 iLoop, pWLoop, &revMask); 3970 }else{ 3971 revMask = pFrom->revLoop; 3972 } 3973 if( isOrdered>=0 && isOrdered<nOrderBy ){ 3974 if( aSortCost[isOrdered]==0 ){ 3975 aSortCost[isOrdered] = whereSortingCost( 3976 pWInfo, nRowEst, nOrderBy, isOrdered 3977 ); 3978 } 3979 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]); 3980 3981 WHERETRACE(0x002, 3982 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 3983 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 3984 rUnsorted, rCost)); 3985 }else{ 3986 rCost = rUnsorted; 3987 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 3988 } 3989 3990 /* Check to see if pWLoop should be added to the set of 3991 ** mxChoice best-so-far paths. 3992 ** 3993 ** First look for an existing path among best-so-far paths 3994 ** that covers the same set of loops and has the same isOrdered 3995 ** setting as the current path candidate. 3996 ** 3997 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 3998 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 3999 ** of legal values for isOrdered, -1..64. 4000 */ 4001 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4002 if( pTo->maskLoop==maskNew 4003 && ((pTo->isOrdered^isOrdered)&0x80)==0 4004 ){ 4005 testcase( jj==nTo-1 ); 4006 break; 4007 } 4008 } 4009 if( jj>=nTo ){ 4010 /* None of the existing best-so-far paths match the candidate. */ 4011 if( nTo>=mxChoice 4012 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4013 ){ 4014 /* The current candidate is no better than any of the mxChoice 4015 ** paths currently in the best-so-far buffer. So discard 4016 ** this candidate as not viable. */ 4017 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4018 if( sqlite3WhereTrace&0x4 ){ 4019 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4020 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4021 isOrdered>=0 ? isOrdered+'0' : '?'); 4022 } 4023 #endif 4024 continue; 4025 } 4026 /* If we reach this points it means that the new candidate path 4027 ** needs to be added to the set of best-so-far paths. */ 4028 if( nTo<mxChoice ){ 4029 /* Increase the size of the aTo set by one */ 4030 jj = nTo++; 4031 }else{ 4032 /* New path replaces the prior worst to keep count below mxChoice */ 4033 jj = mxI; 4034 } 4035 pTo = &aTo[jj]; 4036 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4037 if( sqlite3WhereTrace&0x4 ){ 4038 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4039 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4040 isOrdered>=0 ? isOrdered+'0' : '?'); 4041 } 4042 #endif 4043 }else{ 4044 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4045 ** same set of loops and has the same isOrdered setting as the 4046 ** candidate path. Check to see if the candidate should replace 4047 ** pTo or if the candidate should be skipped. 4048 ** 4049 ** The conditional is an expanded vector comparison equivalent to: 4050 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4051 */ 4052 if( pTo->rCost<rCost 4053 || (pTo->rCost==rCost 4054 && (pTo->nRow<nOut 4055 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4056 ) 4057 ) 4058 ){ 4059 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4060 if( sqlite3WhereTrace&0x4 ){ 4061 sqlite3DebugPrintf( 4062 "Skip %s cost=%-3d,%3d,%3d order=%c", 4063 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4064 isOrdered>=0 ? isOrdered+'0' : '?'); 4065 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4066 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4067 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4068 } 4069 #endif 4070 /* Discard the candidate path from further consideration */ 4071 testcase( pTo->rCost==rCost ); 4072 continue; 4073 } 4074 testcase( pTo->rCost==rCost+1 ); 4075 /* Control reaches here if the candidate path is better than the 4076 ** pTo path. Replace pTo with the candidate. */ 4077 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4078 if( sqlite3WhereTrace&0x4 ){ 4079 sqlite3DebugPrintf( 4080 "Update %s cost=%-3d,%3d,%3d order=%c", 4081 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4082 isOrdered>=0 ? isOrdered+'0' : '?'); 4083 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4084 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4085 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4086 } 4087 #endif 4088 } 4089 /* pWLoop is a winner. Add it to the set of best so far */ 4090 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4091 pTo->revLoop = revMask; 4092 pTo->nRow = nOut; 4093 pTo->rCost = rCost; 4094 pTo->rUnsorted = rUnsorted; 4095 pTo->isOrdered = isOrdered; 4096 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4097 pTo->aLoop[iLoop] = pWLoop; 4098 if( nTo>=mxChoice ){ 4099 mxI = 0; 4100 mxCost = aTo[0].rCost; 4101 mxUnsorted = aTo[0].nRow; 4102 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4103 if( pTo->rCost>mxCost 4104 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4105 ){ 4106 mxCost = pTo->rCost; 4107 mxUnsorted = pTo->rUnsorted; 4108 mxI = jj; 4109 } 4110 } 4111 } 4112 } 4113 } 4114 4115 #ifdef WHERETRACE_ENABLED /* >=2 */ 4116 if( sqlite3WhereTrace & 0x02 ){ 4117 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4118 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4119 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4120 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4121 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4122 if( pTo->isOrdered>0 ){ 4123 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4124 }else{ 4125 sqlite3DebugPrintf("\n"); 4126 } 4127 } 4128 } 4129 #endif 4130 4131 /* Swap the roles of aFrom and aTo for the next generation */ 4132 pFrom = aTo; 4133 aTo = aFrom; 4134 aFrom = pFrom; 4135 nFrom = nTo; 4136 } 4137 4138 if( nFrom==0 ){ 4139 sqlite3ErrorMsg(pParse, "no query solution"); 4140 sqlite3DbFreeNN(db, pSpace); 4141 return SQLITE_ERROR; 4142 } 4143 4144 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4145 pFrom = aFrom; 4146 for(ii=1; ii<nFrom; ii++){ 4147 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4148 } 4149 assert( pWInfo->nLevel==nLoop ); 4150 /* Load the lowest cost path into pWInfo */ 4151 for(iLoop=0; iLoop<nLoop; iLoop++){ 4152 WhereLevel *pLevel = pWInfo->a + iLoop; 4153 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4154 pLevel->iFrom = pWLoop->iTab; 4155 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4156 } 4157 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4158 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4159 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4160 && nRowEst 4161 ){ 4162 Bitmask notUsed; 4163 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4164 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4165 if( rc==pWInfo->pResultSet->nExpr ){ 4166 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4167 } 4168 } 4169 if( pWInfo->pOrderBy ){ 4170 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4171 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4172 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4173 } 4174 }else{ 4175 pWInfo->nOBSat = pFrom->isOrdered; 4176 pWInfo->revMask = pFrom->revLoop; 4177 if( pWInfo->nOBSat<=0 ){ 4178 pWInfo->nOBSat = 0; 4179 if( nLoop>0 ){ 4180 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4181 if( (wsFlags & WHERE_ONEROW)==0 4182 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4183 ){ 4184 Bitmask m = 0; 4185 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4186 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4187 testcase( wsFlags & WHERE_IPK ); 4188 testcase( wsFlags & WHERE_COLUMN_IN ); 4189 if( rc==pWInfo->pOrderBy->nExpr ){ 4190 pWInfo->bOrderedInnerLoop = 1; 4191 pWInfo->revMask = m; 4192 } 4193 } 4194 } 4195 } 4196 } 4197 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4198 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4199 ){ 4200 Bitmask revMask = 0; 4201 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4202 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4203 ); 4204 assert( pWInfo->sorted==0 ); 4205 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4206 pWInfo->sorted = 1; 4207 pWInfo->revMask = revMask; 4208 } 4209 } 4210 } 4211 4212 4213 pWInfo->nRowOut = pFrom->nRow; 4214 4215 /* Free temporary memory and return success */ 4216 sqlite3DbFreeNN(db, pSpace); 4217 return SQLITE_OK; 4218 } 4219 4220 /* 4221 ** Most queries use only a single table (they are not joins) and have 4222 ** simple == constraints against indexed fields. This routine attempts 4223 ** to plan those simple cases using much less ceremony than the 4224 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4225 ** times for the common case. 4226 ** 4227 ** Return non-zero on success, if this query can be handled by this 4228 ** no-frills query planner. Return zero if this query needs the 4229 ** general-purpose query planner. 4230 */ 4231 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4232 WhereInfo *pWInfo; 4233 struct SrcList_item *pItem; 4234 WhereClause *pWC; 4235 WhereTerm *pTerm; 4236 WhereLoop *pLoop; 4237 int iCur; 4238 int j; 4239 Table *pTab; 4240 Index *pIdx; 4241 4242 pWInfo = pBuilder->pWInfo; 4243 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4244 assert( pWInfo->pTabList->nSrc>=1 ); 4245 pItem = pWInfo->pTabList->a; 4246 pTab = pItem->pTab; 4247 if( IsVirtual(pTab) ) return 0; 4248 if( pItem->fg.isIndexedBy ) return 0; 4249 iCur = pItem->iCursor; 4250 pWC = &pWInfo->sWC; 4251 pLoop = pBuilder->pNew; 4252 pLoop->wsFlags = 0; 4253 pLoop->nSkip = 0; 4254 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0); 4255 if( pTerm ){ 4256 testcase( pTerm->eOperator & WO_IS ); 4257 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4258 pLoop->aLTerm[0] = pTerm; 4259 pLoop->nLTerm = 1; 4260 pLoop->u.btree.nEq = 1; 4261 /* TUNING: Cost of a rowid lookup is 10 */ 4262 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 4263 }else{ 4264 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 4265 int opMask; 4266 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 4267 if( !IsUniqueIndex(pIdx) 4268 || pIdx->pPartIdxWhere!=0 4269 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 4270 ) continue; 4271 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 4272 for(j=0; j<pIdx->nKeyCol; j++){ 4273 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx); 4274 if( pTerm==0 ) break; 4275 testcase( pTerm->eOperator & WO_IS ); 4276 pLoop->aLTerm[j] = pTerm; 4277 } 4278 if( j!=pIdx->nKeyCol ) continue; 4279 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 4280 if( pIdx->isCovering || (pItem->colUsed & ~columnsInIndex(pIdx))==0 ){ 4281 pLoop->wsFlags |= WHERE_IDX_ONLY; 4282 } 4283 pLoop->nLTerm = j; 4284 pLoop->u.btree.nEq = j; 4285 pLoop->u.btree.pIndex = pIdx; 4286 /* TUNING: Cost of a unique index lookup is 15 */ 4287 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 4288 break; 4289 } 4290 } 4291 if( pLoop->wsFlags ){ 4292 pLoop->nOut = (LogEst)1; 4293 pWInfo->a[0].pWLoop = pLoop; 4294 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 4295 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 4296 pWInfo->a[0].iTabCur = iCur; 4297 pWInfo->nRowOut = 1; 4298 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 4299 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 4300 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4301 } 4302 #ifdef SQLITE_DEBUG 4303 pLoop->cId = '0'; 4304 #endif 4305 return 1; 4306 } 4307 return 0; 4308 } 4309 4310 /* 4311 ** Helper function for exprIsDeterministic(). 4312 */ 4313 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 4314 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 4315 pWalker->eCode = 0; 4316 return WRC_Abort; 4317 } 4318 return WRC_Continue; 4319 } 4320 4321 /* 4322 ** Return true if the expression contains no non-deterministic SQL 4323 ** functions. Do not consider non-deterministic SQL functions that are 4324 ** part of sub-select statements. 4325 */ 4326 static int exprIsDeterministic(Expr *p){ 4327 Walker w; 4328 memset(&w, 0, sizeof(w)); 4329 w.eCode = 1; 4330 w.xExprCallback = exprNodeIsDeterministic; 4331 sqlite3WalkExpr(&w, p); 4332 return w.eCode; 4333 } 4334 4335 /* 4336 ** Generate the beginning of the loop used for WHERE clause processing. 4337 ** The return value is a pointer to an opaque structure that contains 4338 ** information needed to terminate the loop. Later, the calling routine 4339 ** should invoke sqlite3WhereEnd() with the return value of this function 4340 ** in order to complete the WHERE clause processing. 4341 ** 4342 ** If an error occurs, this routine returns NULL. 4343 ** 4344 ** The basic idea is to do a nested loop, one loop for each table in 4345 ** the FROM clause of a select. (INSERT and UPDATE statements are the 4346 ** same as a SELECT with only a single table in the FROM clause.) For 4347 ** example, if the SQL is this: 4348 ** 4349 ** SELECT * FROM t1, t2, t3 WHERE ...; 4350 ** 4351 ** Then the code generated is conceptually like the following: 4352 ** 4353 ** foreach row1 in t1 do \ Code generated 4354 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 4355 ** foreach row3 in t3 do / 4356 ** ... 4357 ** end \ Code generated 4358 ** end |-- by sqlite3WhereEnd() 4359 ** end / 4360 ** 4361 ** Note that the loops might not be nested in the order in which they 4362 ** appear in the FROM clause if a different order is better able to make 4363 ** use of indices. Note also that when the IN operator appears in 4364 ** the WHERE clause, it might result in additional nested loops for 4365 ** scanning through all values on the right-hand side of the IN. 4366 ** 4367 ** There are Btree cursors associated with each table. t1 uses cursor 4368 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 4369 ** And so forth. This routine generates code to open those VDBE cursors 4370 ** and sqlite3WhereEnd() generates the code to close them. 4371 ** 4372 ** The code that sqlite3WhereBegin() generates leaves the cursors named 4373 ** in pTabList pointing at their appropriate entries. The [...] code 4374 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 4375 ** data from the various tables of the loop. 4376 ** 4377 ** If the WHERE clause is empty, the foreach loops must each scan their 4378 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 4379 ** the tables have indices and there are terms in the WHERE clause that 4380 ** refer to those indices, a complete table scan can be avoided and the 4381 ** code will run much faster. Most of the work of this routine is checking 4382 ** to see if there are indices that can be used to speed up the loop. 4383 ** 4384 ** Terms of the WHERE clause are also used to limit which rows actually 4385 ** make it to the "..." in the middle of the loop. After each "foreach", 4386 ** terms of the WHERE clause that use only terms in that loop and outer 4387 ** loops are evaluated and if false a jump is made around all subsequent 4388 ** inner loops (or around the "..." if the test occurs within the inner- 4389 ** most loop) 4390 ** 4391 ** OUTER JOINS 4392 ** 4393 ** An outer join of tables t1 and t2 is conceptally coded as follows: 4394 ** 4395 ** foreach row1 in t1 do 4396 ** flag = 0 4397 ** foreach row2 in t2 do 4398 ** start: 4399 ** ... 4400 ** flag = 1 4401 ** end 4402 ** if flag==0 then 4403 ** move the row2 cursor to a null row 4404 ** goto start 4405 ** fi 4406 ** end 4407 ** 4408 ** ORDER BY CLAUSE PROCESSING 4409 ** 4410 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 4411 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 4412 ** if there is one. If there is no ORDER BY clause or if this routine 4413 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 4414 ** 4415 ** The iIdxCur parameter is the cursor number of an index. If 4416 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 4417 ** to use for OR clause processing. The WHERE clause should use this 4418 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 4419 ** the first cursor in an array of cursors for all indices. iIdxCur should 4420 ** be used to compute the appropriate cursor depending on which index is 4421 ** used. 4422 */ 4423 WhereInfo *sqlite3WhereBegin( 4424 Parse *pParse, /* The parser context */ 4425 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 4426 Expr *pWhere, /* The WHERE clause */ 4427 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 4428 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 4429 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 4430 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 4431 ** If WHERE_USE_LIMIT, then the limit amount */ 4432 ){ 4433 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4434 int nTabList; /* Number of elements in pTabList */ 4435 WhereInfo *pWInfo; /* Will become the return value of this function */ 4436 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4437 Bitmask notReady; /* Cursors that are not yet positioned */ 4438 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 4439 WhereMaskSet *pMaskSet; /* The expression mask set */ 4440 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 4441 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 4442 int ii; /* Loop counter */ 4443 sqlite3 *db; /* Database connection */ 4444 int rc; /* Return code */ 4445 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 4446 4447 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 4448 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 4449 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4450 )); 4451 4452 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 4453 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4454 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 4455 4456 /* Variable initialization */ 4457 db = pParse->db; 4458 memset(&sWLB, 0, sizeof(sWLB)); 4459 4460 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 4461 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 4462 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 4463 sWLB.pOrderBy = pOrderBy; 4464 4465 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 4466 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 4467 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 4468 wctrlFlags &= ~WHERE_WANT_DISTINCT; 4469 } 4470 4471 /* The number of tables in the FROM clause is limited by the number of 4472 ** bits in a Bitmask 4473 */ 4474 testcase( pTabList->nSrc==BMS ); 4475 if( pTabList->nSrc>BMS ){ 4476 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 4477 return 0; 4478 } 4479 4480 /* This function normally generates a nested loop for all tables in 4481 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 4482 ** only generate code for the first table in pTabList and assume that 4483 ** any cursors associated with subsequent tables are uninitialized. 4484 */ 4485 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 4486 4487 /* Allocate and initialize the WhereInfo structure that will become the 4488 ** return value. A single allocation is used to store the WhereInfo 4489 ** struct, the contents of WhereInfo.a[], the WhereClause structure 4490 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 4491 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 4492 ** some architectures. Hence the ROUND8() below. 4493 */ 4494 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 4495 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 4496 if( db->mallocFailed ){ 4497 sqlite3DbFree(db, pWInfo); 4498 pWInfo = 0; 4499 goto whereBeginError; 4500 } 4501 pWInfo->pParse = pParse; 4502 pWInfo->pTabList = pTabList; 4503 pWInfo->pOrderBy = pOrderBy; 4504 pWInfo->pWhere = pWhere; 4505 pWInfo->pResultSet = pResultSet; 4506 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 4507 pWInfo->nLevel = nTabList; 4508 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(v); 4509 pWInfo->wctrlFlags = wctrlFlags; 4510 pWInfo->iLimit = iAuxArg; 4511 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 4512 memset(&pWInfo->nOBSat, 0, 4513 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 4514 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 4515 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 4516 pMaskSet = &pWInfo->sMaskSet; 4517 sWLB.pWInfo = pWInfo; 4518 sWLB.pWC = &pWInfo->sWC; 4519 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 4520 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 4521 whereLoopInit(sWLB.pNew); 4522 #ifdef SQLITE_DEBUG 4523 sWLB.pNew->cId = '*'; 4524 #endif 4525 4526 /* Split the WHERE clause into separate subexpressions where each 4527 ** subexpression is separated by an AND operator. 4528 */ 4529 initMaskSet(pMaskSet); 4530 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 4531 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 4532 4533 /* Special case: No FROM clause 4534 */ 4535 if( nTabList==0 ){ 4536 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 4537 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4538 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4539 } 4540 } 4541 4542 /* Assign a bit from the bitmask to every term in the FROM clause. 4543 ** 4544 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 4545 ** 4546 ** The rule of the previous sentence ensures thta if X is the bitmask for 4547 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 4548 ** Knowing the bitmask for all tables to the left of a left join is 4549 ** important. Ticket #3015. 4550 ** 4551 ** Note that bitmasks are created for all pTabList->nSrc tables in 4552 ** pTabList, not just the first nTabList tables. nTabList is normally 4553 ** equal to pTabList->nSrc but might be shortened to 1 if the 4554 ** WHERE_OR_SUBCLAUSE flag is set. 4555 */ 4556 for(ii=0; ii<pTabList->nSrc; ii++){ 4557 createMask(pMaskSet, pTabList->a[ii].iCursor); 4558 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 4559 } 4560 #ifdef SQLITE_DEBUG 4561 { 4562 Bitmask mx = 0; 4563 for(ii=0; ii<pTabList->nSrc; ii++){ 4564 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 4565 assert( m>=mx ); 4566 mx = m; 4567 } 4568 } 4569 #endif 4570 4571 /* Analyze all of the subexpressions. */ 4572 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 4573 if( db->mallocFailed ) goto whereBeginError; 4574 4575 /* Special case: WHERE terms that do not refer to any tables in the join 4576 ** (constant expressions). Evaluate each such term, and jump over all the 4577 ** generated code if the result is not true. 4578 ** 4579 ** Do not do this if the expression contains non-deterministic functions 4580 ** that are not within a sub-select. This is not strictly required, but 4581 ** preserves SQLite's legacy behaviour in the following two cases: 4582 ** 4583 ** FROM ... WHERE random()>0; -- eval random() once per row 4584 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 4585 */ 4586 for(ii=0; ii<sWLB.pWC->nTerm; ii++){ 4587 WhereTerm *pT = &sWLB.pWC->a[ii]; 4588 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 4589 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 4590 pT->wtFlags |= TERM_CODED; 4591 } 4592 } 4593 4594 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4595 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 4596 /* The DISTINCT marking is pointless. Ignore it. */ 4597 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4598 }else if( pOrderBy==0 ){ 4599 /* Try to ORDER BY the result set to make distinct processing easier */ 4600 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 4601 pWInfo->pOrderBy = pResultSet; 4602 } 4603 } 4604 4605 /* Construct the WhereLoop objects */ 4606 #if defined(WHERETRACE_ENABLED) 4607 if( sqlite3WhereTrace & 0xffff ){ 4608 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 4609 if( wctrlFlags & WHERE_USE_LIMIT ){ 4610 sqlite3DebugPrintf(", limit: %d", iAuxArg); 4611 } 4612 sqlite3DebugPrintf(")\n"); 4613 } 4614 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 4615 sqlite3WhereClausePrint(sWLB.pWC); 4616 } 4617 #endif 4618 4619 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 4620 rc = whereLoopAddAll(&sWLB); 4621 if( rc ) goto whereBeginError; 4622 4623 #ifdef WHERETRACE_ENABLED 4624 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4625 WhereLoop *p; 4626 int i; 4627 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4628 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4629 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4630 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 4631 whereLoopPrint(p, sWLB.pWC); 4632 } 4633 } 4634 #endif 4635 4636 wherePathSolver(pWInfo, 0); 4637 if( db->mallocFailed ) goto whereBeginError; 4638 if( pWInfo->pOrderBy ){ 4639 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 4640 if( db->mallocFailed ) goto whereBeginError; 4641 } 4642 } 4643 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 4644 pWInfo->revMask = ALLBITS; 4645 } 4646 if( pParse->nErr || NEVER(db->mallocFailed) ){ 4647 goto whereBeginError; 4648 } 4649 #ifdef WHERETRACE_ENABLED 4650 if( sqlite3WhereTrace ){ 4651 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 4652 if( pWInfo->nOBSat>0 ){ 4653 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 4654 } 4655 switch( pWInfo->eDistinct ){ 4656 case WHERE_DISTINCT_UNIQUE: { 4657 sqlite3DebugPrintf(" DISTINCT=unique"); 4658 break; 4659 } 4660 case WHERE_DISTINCT_ORDERED: { 4661 sqlite3DebugPrintf(" DISTINCT=ordered"); 4662 break; 4663 } 4664 case WHERE_DISTINCT_UNORDERED: { 4665 sqlite3DebugPrintf(" DISTINCT=unordered"); 4666 break; 4667 } 4668 } 4669 sqlite3DebugPrintf("\n"); 4670 for(ii=0; ii<pWInfo->nLevel; ii++){ 4671 whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 4672 } 4673 } 4674 #endif 4675 /* Attempt to omit tables from the join that do not effect the result */ 4676 if( pWInfo->nLevel>=2 4677 && pResultSet!=0 4678 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 4679 ){ 4680 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet); 4681 if( sWLB.pOrderBy ){ 4682 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy); 4683 } 4684 while( pWInfo->nLevel>=2 ){ 4685 WhereTerm *pTerm, *pEnd; 4686 pLoop = pWInfo->a[pWInfo->nLevel-1].pWLoop; 4687 if( (pWInfo->pTabList->a[pLoop->iTab].fg.jointype & JT_LEFT)==0 ) break; 4688 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 4689 && (pLoop->wsFlags & WHERE_ONEROW)==0 4690 ){ 4691 break; 4692 } 4693 if( (tabUsed & pLoop->maskSelf)!=0 ) break; 4694 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; 4695 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4696 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 4697 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 4698 ){ 4699 break; 4700 } 4701 } 4702 if( pTerm<pEnd ) break; 4703 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 4704 pWInfo->nLevel--; 4705 nTabList--; 4706 } 4707 } 4708 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 4709 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 4710 4711 /* If the caller is an UPDATE or DELETE statement that is requesting 4712 ** to use a one-pass algorithm, determine if this is appropriate. 4713 */ 4714 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 4715 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 4716 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 4717 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 4718 if( bOnerow 4719 || ((wctrlFlags & WHERE_ONEPASS_MULTIROW)!=0 4720 && 0==(wsFlags & WHERE_VIRTUALTABLE)) 4721 ){ 4722 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 4723 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 4724 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 4725 bFordelete = OPFLAG_FORDELETE; 4726 } 4727 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 4728 } 4729 } 4730 } 4731 4732 /* Open all tables in the pTabList and any indices selected for 4733 ** searching those tables. 4734 */ 4735 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 4736 Table *pTab; /* Table to open */ 4737 int iDb; /* Index of database containing table/index */ 4738 struct SrcList_item *pTabItem; 4739 4740 pTabItem = &pTabList->a[pLevel->iFrom]; 4741 pTab = pTabItem->pTab; 4742 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 4743 pLoop = pLevel->pWLoop; 4744 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 4745 /* Do nothing */ 4746 }else 4747 #ifndef SQLITE_OMIT_VIRTUALTABLE 4748 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 4749 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 4750 int iCur = pTabItem->iCursor; 4751 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 4752 }else if( IsVirtual(pTab) ){ 4753 /* noop */ 4754 }else 4755 #endif 4756 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 4757 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 4758 int op = OP_OpenRead; 4759 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4760 op = OP_OpenWrite; 4761 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 4762 }; 4763 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 4764 assert( pTabItem->iCursor==pLevel->iTabCur ); 4765 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 4766 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 4767 if( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol<BMS && HasRowid(pTab) ){ 4768 Bitmask b = pTabItem->colUsed; 4769 int n = 0; 4770 for(; b; b=b>>1, n++){} 4771 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 4772 assert( n<=pTab->nCol ); 4773 } 4774 #ifdef SQLITE_ENABLE_CURSOR_HINTS 4775 if( pLoop->u.btree.pIndex!=0 ){ 4776 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 4777 }else 4778 #endif 4779 { 4780 sqlite3VdbeChangeP5(v, bFordelete); 4781 } 4782 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 4783 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 4784 (const u8*)&pTabItem->colUsed, P4_INT64); 4785 #endif 4786 }else{ 4787 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 4788 } 4789 if( pLoop->wsFlags & WHERE_INDEXED ){ 4790 Index *pIx = pLoop->u.btree.pIndex; 4791 int iIndexCur; 4792 int op = OP_OpenRead; 4793 /* iAuxArg is always set if to a positive value if ONEPASS is possible */ 4794 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 4795 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 4796 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 4797 ){ 4798 /* This is one term of an OR-optimization using the PRIMARY KEY of a 4799 ** WITHOUT ROWID table. No need for a separate index */ 4800 iIndexCur = pLevel->iTabCur; 4801 op = 0; 4802 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4803 Index *pJ = pTabItem->pTab->pIndex; 4804 iIndexCur = iAuxArg; 4805 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 4806 while( ALWAYS(pJ) && pJ!=pIx ){ 4807 iIndexCur++; 4808 pJ = pJ->pNext; 4809 } 4810 op = OP_OpenWrite; 4811 pWInfo->aiCurOnePass[1] = iIndexCur; 4812 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 4813 iIndexCur = iAuxArg; 4814 op = OP_ReopenIdx; 4815 }else{ 4816 iIndexCur = pParse->nTab++; 4817 } 4818 pLevel->iIdxCur = iIndexCur; 4819 assert( pIx->pSchema==pTab->pSchema ); 4820 assert( iIndexCur>=0 ); 4821 if( op ){ 4822 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 4823 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 4824 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 4825 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 4826 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 4827 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 4828 ){ 4829 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); /* Hint to COMDB2 */ 4830 } 4831 VdbeComment((v, "%s", pIx->zName)); 4832 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 4833 { 4834 u64 colUsed = 0; 4835 int ii, jj; 4836 for(ii=0; ii<pIx->nColumn; ii++){ 4837 jj = pIx->aiColumn[ii]; 4838 if( jj<0 ) continue; 4839 if( jj>63 ) jj = 63; 4840 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 4841 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 4842 } 4843 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 4844 (u8*)&colUsed, P4_INT64); 4845 } 4846 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 4847 } 4848 } 4849 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 4850 } 4851 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 4852 if( db->mallocFailed ) goto whereBeginError; 4853 4854 /* Generate the code to do the search. Each iteration of the for 4855 ** loop below generates code for a single nested loop of the VM 4856 ** program. 4857 */ 4858 notReady = ~(Bitmask)0; 4859 for(ii=0; ii<nTabList; ii++){ 4860 int addrExplain; 4861 int wsFlags; 4862 pLevel = &pWInfo->a[ii]; 4863 wsFlags = pLevel->pWLoop->wsFlags; 4864 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 4865 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 4866 constructAutomaticIndex(pParse, &pWInfo->sWC, 4867 &pTabList->a[pLevel->iFrom], notReady, pLevel); 4868 if( db->mallocFailed ) goto whereBeginError; 4869 } 4870 #endif 4871 addrExplain = sqlite3WhereExplainOneScan( 4872 pParse, pTabList, pLevel, ii, pLevel->iFrom, wctrlFlags 4873 ); 4874 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 4875 notReady = sqlite3WhereCodeOneLoopStart(pWInfo, ii, notReady); 4876 pWInfo->iContinue = pLevel->addrCont; 4877 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 4878 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 4879 } 4880 } 4881 4882 /* Done. */ 4883 VdbeModuleComment((v, "Begin WHERE-core")); 4884 return pWInfo; 4885 4886 /* Jump here if malloc fails */ 4887 whereBeginError: 4888 if( pWInfo ){ 4889 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 4890 whereInfoFree(db, pWInfo); 4891 } 4892 return 0; 4893 } 4894 4895 /* 4896 ** Generate the end of the WHERE loop. See comments on 4897 ** sqlite3WhereBegin() for additional information. 4898 */ 4899 void sqlite3WhereEnd(WhereInfo *pWInfo){ 4900 Parse *pParse = pWInfo->pParse; 4901 Vdbe *v = pParse->pVdbe; 4902 int i; 4903 WhereLevel *pLevel; 4904 WhereLoop *pLoop; 4905 SrcList *pTabList = pWInfo->pTabList; 4906 sqlite3 *db = pParse->db; 4907 4908 /* Generate loop termination code. 4909 */ 4910 VdbeModuleComment((v, "End WHERE-core")); 4911 sqlite3ExprCacheClear(pParse); 4912 for(i=pWInfo->nLevel-1; i>=0; i--){ 4913 int addr; 4914 pLevel = &pWInfo->a[i]; 4915 pLoop = pLevel->pWLoop; 4916 if( pLevel->op!=OP_Noop ){ 4917 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 4918 int addrSeek = 0; 4919 Index *pIdx; 4920 int n; 4921 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 4922 && (pLoop->wsFlags & WHERE_INDEXED)!=0 4923 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 4924 && (n = pLoop->u.btree.nIdxCol)>0 4925 && pIdx->aiRowLogEst[n]>=36 4926 ){ 4927 int r1 = pParse->nMem+1; 4928 int j, op; 4929 for(j=0; j<n; j++){ 4930 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 4931 } 4932 pParse->nMem += n+1; 4933 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 4934 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 4935 VdbeCoverageIf(v, op==OP_SeekLT); 4936 VdbeCoverageIf(v, op==OP_SeekGT); 4937 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 4938 } 4939 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 4940 /* The common case: Advance to the next row */ 4941 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 4942 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 4943 sqlite3VdbeChangeP5(v, pLevel->p5); 4944 VdbeCoverage(v); 4945 VdbeCoverageIf(v, pLevel->op==OP_Next); 4946 VdbeCoverageIf(v, pLevel->op==OP_Prev); 4947 VdbeCoverageIf(v, pLevel->op==OP_VNext); 4948 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 4949 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 4950 #endif 4951 }else{ 4952 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 4953 } 4954 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 4955 struct InLoop *pIn; 4956 int j; 4957 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 4958 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 4959 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 4960 if( pIn->eEndLoopOp!=OP_Noop ){ 4961 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 4962 VdbeCoverage(v); 4963 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_PrevIfOpen); 4964 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_NextIfOpen); 4965 } 4966 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 4967 } 4968 } 4969 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 4970 if( pLevel->addrSkip ){ 4971 sqlite3VdbeGoto(v, pLevel->addrSkip); 4972 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 4973 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 4974 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 4975 } 4976 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 4977 if( pLevel->addrLikeRep ){ 4978 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 4979 pLevel->addrLikeRep); 4980 VdbeCoverage(v); 4981 } 4982 #endif 4983 if( pLevel->iLeftJoin ){ 4984 int ws = pLoop->wsFlags; 4985 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 4986 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 4987 if( (ws & WHERE_IDX_ONLY)==0 ){ 4988 sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor); 4989 } 4990 if( (ws & WHERE_INDEXED) 4991 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCovidx) 4992 ){ 4993 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 4994 } 4995 if( pLevel->op==OP_Return ){ 4996 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 4997 }else{ 4998 sqlite3VdbeGoto(v, pLevel->addrFirst); 4999 } 5000 sqlite3VdbeJumpHere(v, addr); 5001 } 5002 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 5003 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 5004 } 5005 5006 /* The "break" point is here, just past the end of the outer loop. 5007 ** Set it. 5008 */ 5009 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5010 5011 assert( pWInfo->nLevel<=pTabList->nSrc ); 5012 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5013 int k, last; 5014 VdbeOp *pOp; 5015 Index *pIdx = 0; 5016 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; 5017 Table *pTab = pTabItem->pTab; 5018 assert( pTab!=0 ); 5019 pLoop = pLevel->pWLoop; 5020 5021 /* For a co-routine, change all OP_Column references to the table of 5022 ** the co-routine into OP_Copy of result contained in a register. 5023 ** OP_Rowid becomes OP_Null. 5024 */ 5025 if( pTabItem->fg.viaCoroutine ){ 5026 testcase( pParse->db->mallocFailed ); 5027 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 5028 pTabItem->regResult, 0); 5029 continue; 5030 } 5031 5032 /* If this scan uses an index, make VDBE code substitutions to read data 5033 ** from the index instead of from the table where possible. In some cases 5034 ** this optimization prevents the table from ever being read, which can 5035 ** yield a significant performance boost. 5036 ** 5037 ** Calls to the code generator in between sqlite3WhereBegin and 5038 ** sqlite3WhereEnd will have created code that references the table 5039 ** directly. This loop scans all that code looking for opcodes 5040 ** that reference the table and converts them into opcodes that 5041 ** reference the index. 5042 */ 5043 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 5044 pIdx = pLoop->u.btree.pIndex; 5045 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 5046 pIdx = pLevel->u.pCovidx; 5047 } 5048 if( pIdx 5049 && (pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable)) 5050 && !db->mallocFailed 5051 ){ 5052 last = sqlite3VdbeCurrentAddr(v); 5053 k = pLevel->addrBody; 5054 pOp = sqlite3VdbeGetOp(v, k); 5055 for(; k<last; k++, pOp++){ 5056 if( pOp->p1!=pLevel->iTabCur ) continue; 5057 if( pOp->opcode==OP_Column ){ 5058 int x = pOp->p2; 5059 assert( pIdx->pTable==pTab ); 5060 if( !HasRowid(pTab) ){ 5061 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 5062 x = pPk->aiColumn[x]; 5063 assert( x>=0 ); 5064 } 5065 x = sqlite3ColumnOfIndex(pIdx, x); 5066 if( x>=0 ){ 5067 pOp->p2 = x; 5068 pOp->p1 = pLevel->iIdxCur; 5069 } 5070 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 5071 || pWInfo->eOnePass ); 5072 }else if( pOp->opcode==OP_Rowid ){ 5073 pOp->p1 = pLevel->iIdxCur; 5074 pOp->opcode = OP_IdxRowid; 5075 }else if( pOp->opcode==OP_IfNullRow ){ 5076 pOp->p1 = pLevel->iIdxCur; 5077 } 5078 } 5079 } 5080 } 5081 5082 /* Final cleanup 5083 */ 5084 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5085 whereInfoFree(db, pWInfo); 5086 return; 5087 } 5088