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_EQUIV))==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 u16 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_EQUIV))==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 = pTerm->eOperator & WO_ALL; 933 if( op==WO_IN ) op = WO_EQ; 934 if( op==WO_AUX ){ 935 pIdxCons[j].op = pTerm->eMatchOp; 936 }else if( op & (WO_ISNULL|WO_IS) ){ 937 if( op==WO_ISNULL ){ 938 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_ISNULL; 939 }else{ 940 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_IS; 941 } 942 }else{ 943 pIdxCons[j].op = (u8)op; 944 /* The direct assignment in the previous line is possible only because 945 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 946 ** following asserts verify this fact. */ 947 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 948 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 949 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 950 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 951 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 952 assert( pTerm->eOperator&(WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_AUX) ); 953 954 if( op & (WO_LT|WO_LE|WO_GT|WO_GE) 955 && sqlite3ExprIsVector(pTerm->pExpr->pRight) 956 ){ 957 if( i<16 ) mNoOmit |= (1 << i); 958 if( op==WO_LT ) pIdxCons[j].op = WO_LE; 959 if( op==WO_GT ) pIdxCons[j].op = WO_GE; 960 } 961 } 962 963 j++; 964 } 965 for(i=0; i<nOrderBy; i++){ 966 Expr *pExpr = pOrderBy->a[i].pExpr; 967 pIdxOrderBy[i].iColumn = pExpr->iColumn; 968 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder; 969 } 970 971 *pmNoOmit = mNoOmit; 972 return pIdxInfo; 973 } 974 975 /* 976 ** The table object reference passed as the second argument to this function 977 ** must represent a virtual table. This function invokes the xBestIndex() 978 ** method of the virtual table with the sqlite3_index_info object that 979 ** comes in as the 3rd argument to this function. 980 ** 981 ** If an error occurs, pParse is populated with an error message and a 982 ** non-zero value is returned. Otherwise, 0 is returned and the output 983 ** part of the sqlite3_index_info structure is left populated. 984 ** 985 ** Whether or not an error is returned, it is the responsibility of the 986 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 987 ** that this is required. 988 */ 989 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 990 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 991 int rc; 992 993 TRACE_IDX_INPUTS(p); 994 rc = pVtab->pModule->xBestIndex(pVtab, p); 995 TRACE_IDX_OUTPUTS(p); 996 997 if( rc!=SQLITE_OK ){ 998 if( rc==SQLITE_NOMEM ){ 999 sqlite3OomFault(pParse->db); 1000 }else if( !pVtab->zErrMsg ){ 1001 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 1002 }else{ 1003 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 1004 } 1005 } 1006 sqlite3_free(pVtab->zErrMsg); 1007 pVtab->zErrMsg = 0; 1008 1009 #if 0 1010 /* This error is now caught by the caller. 1011 ** Search for "xBestIndex malfunction" below */ 1012 for(i=0; i<p->nConstraint; i++){ 1013 if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){ 1014 sqlite3ErrorMsg(pParse, 1015 "table %s: xBestIndex returned an invalid plan", pTab->zName); 1016 } 1017 } 1018 #endif 1019 1020 return pParse->nErr; 1021 } 1022 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 1023 1024 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1025 /* 1026 ** Estimate the location of a particular key among all keys in an 1027 ** index. Store the results in aStat as follows: 1028 ** 1029 ** aStat[0] Est. number of rows less than pRec 1030 ** aStat[1] Est. number of rows equal to pRec 1031 ** 1032 ** Return the index of the sample that is the smallest sample that 1033 ** is greater than or equal to pRec. Note that this index is not an index 1034 ** into the aSample[] array - it is an index into a virtual set of samples 1035 ** based on the contents of aSample[] and the number of fields in record 1036 ** pRec. 1037 */ 1038 static int whereKeyStats( 1039 Parse *pParse, /* Database connection */ 1040 Index *pIdx, /* Index to consider domain of */ 1041 UnpackedRecord *pRec, /* Vector of values to consider */ 1042 int roundUp, /* Round up if true. Round down if false */ 1043 tRowcnt *aStat /* OUT: stats written here */ 1044 ){ 1045 IndexSample *aSample = pIdx->aSample; 1046 int iCol; /* Index of required stats in anEq[] etc. */ 1047 int i; /* Index of first sample >= pRec */ 1048 int iSample; /* Smallest sample larger than or equal to pRec */ 1049 int iMin = 0; /* Smallest sample not yet tested */ 1050 int iTest; /* Next sample to test */ 1051 int res; /* Result of comparison operation */ 1052 int nField; /* Number of fields in pRec */ 1053 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 1054 1055 #ifndef SQLITE_DEBUG 1056 UNUSED_PARAMETER( pParse ); 1057 #endif 1058 assert( pRec!=0 ); 1059 assert( pIdx->nSample>0 ); 1060 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1061 1062 /* Do a binary search to find the first sample greater than or equal 1063 ** to pRec. If pRec contains a single field, the set of samples to search 1064 ** is simply the aSample[] array. If the samples in aSample[] contain more 1065 ** than one fields, all fields following the first are ignored. 1066 ** 1067 ** If pRec contains N fields, where N is more than one, then as well as the 1068 ** samples in aSample[] (truncated to N fields), the search also has to 1069 ** consider prefixes of those samples. For example, if the set of samples 1070 ** in aSample is: 1071 ** 1072 ** aSample[0] = (a, 5) 1073 ** aSample[1] = (a, 10) 1074 ** aSample[2] = (b, 5) 1075 ** aSample[3] = (c, 100) 1076 ** aSample[4] = (c, 105) 1077 ** 1078 ** Then the search space should ideally be the samples above and the 1079 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1080 ** the code actually searches this set: 1081 ** 1082 ** 0: (a) 1083 ** 1: (a, 5) 1084 ** 2: (a, 10) 1085 ** 3: (a, 10) 1086 ** 4: (b) 1087 ** 5: (b, 5) 1088 ** 6: (c) 1089 ** 7: (c, 100) 1090 ** 8: (c, 105) 1091 ** 9: (c, 105) 1092 ** 1093 ** For each sample in the aSample[] array, N samples are present in the 1094 ** effective sample array. In the above, samples 0 and 1 are based on 1095 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1096 ** 1097 ** Often, sample i of each block of N effective samples has (i+1) fields. 1098 ** Except, each sample may be extended to ensure that it is greater than or 1099 ** equal to the previous sample in the array. For example, in the above, 1100 ** sample 2 is the first sample of a block of N samples, so at first it 1101 ** appears that it should be 1 field in size. However, that would make it 1102 ** smaller than sample 1, so the binary search would not work. As a result, 1103 ** it is extended to two fields. The duplicates that this creates do not 1104 ** cause any problems. 1105 */ 1106 nField = pRec->nField; 1107 iCol = 0; 1108 iSample = pIdx->nSample * nField; 1109 do{ 1110 int iSamp; /* Index in aSample[] of test sample */ 1111 int n; /* Number of fields in test sample */ 1112 1113 iTest = (iMin+iSample)/2; 1114 iSamp = iTest / nField; 1115 if( iSamp>0 ){ 1116 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1117 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1118 ** fields that is greater than the previous effective sample. */ 1119 for(n=(iTest % nField) + 1; n<nField; n++){ 1120 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1121 } 1122 }else{ 1123 n = iTest + 1; 1124 } 1125 1126 pRec->nField = n; 1127 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1128 if( res<0 ){ 1129 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1130 iMin = iTest+1; 1131 }else if( res==0 && n<nField ){ 1132 iLower = aSample[iSamp].anLt[n-1]; 1133 iMin = iTest+1; 1134 res = -1; 1135 }else{ 1136 iSample = iTest; 1137 iCol = n-1; 1138 } 1139 }while( res && iMin<iSample ); 1140 i = iSample / nField; 1141 1142 #ifdef SQLITE_DEBUG 1143 /* The following assert statements check that the binary search code 1144 ** above found the right answer. This block serves no purpose other 1145 ** than to invoke the asserts. */ 1146 if( pParse->db->mallocFailed==0 ){ 1147 if( res==0 ){ 1148 /* If (res==0) is true, then pRec must be equal to sample i. */ 1149 assert( i<pIdx->nSample ); 1150 assert( iCol==nField-1 ); 1151 pRec->nField = nField; 1152 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1153 || pParse->db->mallocFailed 1154 ); 1155 }else{ 1156 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1157 ** all samples in the aSample[] array, pRec must be smaller than the 1158 ** (iCol+1) field prefix of sample i. */ 1159 assert( i<=pIdx->nSample && i>=0 ); 1160 pRec->nField = iCol+1; 1161 assert( i==pIdx->nSample 1162 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1163 || pParse->db->mallocFailed ); 1164 1165 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1166 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1167 ** be greater than or equal to the (iCol) field prefix of sample i. 1168 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1169 if( iCol>0 ){ 1170 pRec->nField = iCol; 1171 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1172 || pParse->db->mallocFailed ); 1173 } 1174 if( i>0 ){ 1175 pRec->nField = nField; 1176 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1177 || pParse->db->mallocFailed ); 1178 } 1179 } 1180 } 1181 #endif /* ifdef SQLITE_DEBUG */ 1182 1183 if( res==0 ){ 1184 /* Record pRec is equal to sample i */ 1185 assert( iCol==nField-1 ); 1186 aStat[0] = aSample[i].anLt[iCol]; 1187 aStat[1] = aSample[i].anEq[iCol]; 1188 }else{ 1189 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1190 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1191 ** is larger than all samples in the array. */ 1192 tRowcnt iUpper, iGap; 1193 if( i>=pIdx->nSample ){ 1194 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1195 }else{ 1196 iUpper = aSample[i].anLt[iCol]; 1197 } 1198 1199 if( iLower>=iUpper ){ 1200 iGap = 0; 1201 }else{ 1202 iGap = iUpper - iLower; 1203 } 1204 if( roundUp ){ 1205 iGap = (iGap*2)/3; 1206 }else{ 1207 iGap = iGap/3; 1208 } 1209 aStat[0] = iLower + iGap; 1210 aStat[1] = pIdx->aAvgEq[nField-1]; 1211 } 1212 1213 /* Restore the pRec->nField value before returning. */ 1214 pRec->nField = nField; 1215 return i; 1216 } 1217 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1218 1219 /* 1220 ** If it is not NULL, pTerm is a term that provides an upper or lower 1221 ** bound on a range scan. Without considering pTerm, it is estimated 1222 ** that the scan will visit nNew rows. This function returns the number 1223 ** estimated to be visited after taking pTerm into account. 1224 ** 1225 ** If the user explicitly specified a likelihood() value for this term, 1226 ** then the return value is the likelihood multiplied by the number of 1227 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1228 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1229 */ 1230 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1231 LogEst nRet = nNew; 1232 if( pTerm ){ 1233 if( pTerm->truthProb<=0 ){ 1234 nRet += pTerm->truthProb; 1235 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1236 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1237 } 1238 } 1239 return nRet; 1240 } 1241 1242 1243 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1244 /* 1245 ** Return the affinity for a single column of an index. 1246 */ 1247 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1248 assert( iCol>=0 && iCol<pIdx->nColumn ); 1249 if( !pIdx->zColAff ){ 1250 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1251 } 1252 return pIdx->zColAff[iCol]; 1253 } 1254 #endif 1255 1256 1257 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1258 /* 1259 ** This function is called to estimate the number of rows visited by a 1260 ** range-scan on a skip-scan index. For example: 1261 ** 1262 ** CREATE INDEX i1 ON t1(a, b, c); 1263 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1264 ** 1265 ** Value pLoop->nOut is currently set to the estimated number of rows 1266 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1267 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1268 ** on the stat4 data for the index. this scan will be peformed multiple 1269 ** times (once for each (a,b) combination that matches a=?) is dealt with 1270 ** by the caller. 1271 ** 1272 ** It does this by scanning through all stat4 samples, comparing values 1273 ** extracted from pLower and pUpper with the corresponding column in each 1274 ** sample. If L and U are the number of samples found to be less than or 1275 ** equal to the values extracted from pLower and pUpper respectively, and 1276 ** N is the total number of samples, the pLoop->nOut value is adjusted 1277 ** as follows: 1278 ** 1279 ** nOut = nOut * ( min(U - L, 1) / N ) 1280 ** 1281 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1282 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1283 ** U is set to N. 1284 ** 1285 ** Normally, this function sets *pbDone to 1 before returning. However, 1286 ** if no value can be extracted from either pLower or pUpper (and so the 1287 ** estimate of the number of rows delivered remains unchanged), *pbDone 1288 ** is left as is. 1289 ** 1290 ** If an error occurs, an SQLite error code is returned. Otherwise, 1291 ** SQLITE_OK. 1292 */ 1293 static int whereRangeSkipScanEst( 1294 Parse *pParse, /* Parsing & code generating context */ 1295 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1296 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1297 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1298 int *pbDone /* Set to true if at least one expr. value extracted */ 1299 ){ 1300 Index *p = pLoop->u.btree.pIndex; 1301 int nEq = pLoop->u.btree.nEq; 1302 sqlite3 *db = pParse->db; 1303 int nLower = -1; 1304 int nUpper = p->nSample+1; 1305 int rc = SQLITE_OK; 1306 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1307 CollSeq *pColl; 1308 1309 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1310 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1311 sqlite3_value *pVal = 0; /* Value extracted from record */ 1312 1313 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1314 if( pLower ){ 1315 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1316 nLower = 0; 1317 } 1318 if( pUpper && rc==SQLITE_OK ){ 1319 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1320 nUpper = p2 ? 0 : p->nSample; 1321 } 1322 1323 if( p1 || p2 ){ 1324 int i; 1325 int nDiff; 1326 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1327 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1328 if( rc==SQLITE_OK && p1 ){ 1329 int res = sqlite3MemCompare(p1, pVal, pColl); 1330 if( res>=0 ) nLower++; 1331 } 1332 if( rc==SQLITE_OK && p2 ){ 1333 int res = sqlite3MemCompare(p2, pVal, pColl); 1334 if( res>=0 ) nUpper++; 1335 } 1336 } 1337 nDiff = (nUpper - nLower); 1338 if( nDiff<=0 ) nDiff = 1; 1339 1340 /* If there is both an upper and lower bound specified, and the 1341 ** comparisons indicate that they are close together, use the fallback 1342 ** method (assume that the scan visits 1/64 of the rows) for estimating 1343 ** the number of rows visited. Otherwise, estimate the number of rows 1344 ** using the method described in the header comment for this function. */ 1345 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1346 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1347 pLoop->nOut -= nAdjust; 1348 *pbDone = 1; 1349 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1350 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1351 } 1352 1353 }else{ 1354 assert( *pbDone==0 ); 1355 } 1356 1357 sqlite3ValueFree(p1); 1358 sqlite3ValueFree(p2); 1359 sqlite3ValueFree(pVal); 1360 1361 return rc; 1362 } 1363 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1364 1365 /* 1366 ** This function is used to estimate the number of rows that will be visited 1367 ** by scanning an index for a range of values. The range may have an upper 1368 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1369 ** and lower bounds are represented by pLower and pUpper respectively. For 1370 ** example, assuming that index p is on t1(a): 1371 ** 1372 ** ... FROM t1 WHERE a > ? AND a < ? ... 1373 ** |_____| |_____| 1374 ** | | 1375 ** pLower pUpper 1376 ** 1377 ** If either of the upper or lower bound is not present, then NULL is passed in 1378 ** place of the corresponding WhereTerm. 1379 ** 1380 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1381 ** column subject to the range constraint. Or, equivalently, the number of 1382 ** equality constraints optimized by the proposed index scan. For example, 1383 ** assuming index p is on t1(a, b), and the SQL query is: 1384 ** 1385 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1386 ** 1387 ** then nEq is set to 1 (as the range restricted column, b, is the second 1388 ** left-most column of the index). Or, if the query is: 1389 ** 1390 ** ... FROM t1 WHERE a > ? AND a < ? ... 1391 ** 1392 ** then nEq is set to 0. 1393 ** 1394 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1395 ** number of rows that the index scan is expected to visit without 1396 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1397 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1398 ** to account for the range constraints pLower and pUpper. 1399 ** 1400 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1401 ** used, a single range inequality reduces the search space by a factor of 4. 1402 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1403 ** rows visited by a factor of 64. 1404 */ 1405 static int whereRangeScanEst( 1406 Parse *pParse, /* Parsing & code generating context */ 1407 WhereLoopBuilder *pBuilder, 1408 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1409 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1410 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1411 ){ 1412 int rc = SQLITE_OK; 1413 int nOut = pLoop->nOut; 1414 LogEst nNew; 1415 1416 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1417 Index *p = pLoop->u.btree.pIndex; 1418 int nEq = pLoop->u.btree.nEq; 1419 1420 if( p->nSample>0 && nEq<p->nSampleCol ){ 1421 if( nEq==pBuilder->nRecValid ){ 1422 UnpackedRecord *pRec = pBuilder->pRec; 1423 tRowcnt a[2]; 1424 int nBtm = pLoop->u.btree.nBtm; 1425 int nTop = pLoop->u.btree.nTop; 1426 1427 /* Variable iLower will be set to the estimate of the number of rows in 1428 ** the index that are less than the lower bound of the range query. The 1429 ** lower bound being the concatenation of $P and $L, where $P is the 1430 ** key-prefix formed by the nEq values matched against the nEq left-most 1431 ** columns of the index, and $L is the value in pLower. 1432 ** 1433 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1434 ** is not a simple variable or literal value), the lower bound of the 1435 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1436 ** if $L is available, whereKeyStats() is called for both ($P) and 1437 ** ($P:$L) and the larger of the two returned values is used. 1438 ** 1439 ** Similarly, iUpper is to be set to the estimate of the number of rows 1440 ** less than the upper bound of the range query. Where the upper bound 1441 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1442 ** of iUpper are requested of whereKeyStats() and the smaller used. 1443 ** 1444 ** The number of rows between the two bounds is then just iUpper-iLower. 1445 */ 1446 tRowcnt iLower; /* Rows less than the lower bound */ 1447 tRowcnt iUpper; /* Rows less than the upper bound */ 1448 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1449 int iUprIdx = -1; /* aSample[] for the upper bound */ 1450 1451 if( pRec ){ 1452 testcase( pRec->nField!=pBuilder->nRecValid ); 1453 pRec->nField = pBuilder->nRecValid; 1454 } 1455 /* Determine iLower and iUpper using ($P) only. */ 1456 if( nEq==0 ){ 1457 iLower = 0; 1458 iUpper = p->nRowEst0; 1459 }else{ 1460 /* Note: this call could be optimized away - since the same values must 1461 ** have been requested when testing key $P in whereEqualScanEst(). */ 1462 whereKeyStats(pParse, p, pRec, 0, a); 1463 iLower = a[0]; 1464 iUpper = a[0] + a[1]; 1465 } 1466 1467 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1468 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1469 assert( p->aSortOrder!=0 ); 1470 if( p->aSortOrder[nEq] ){ 1471 /* The roles of pLower and pUpper are swapped for a DESC index */ 1472 SWAP(WhereTerm*, pLower, pUpper); 1473 SWAP(int, nBtm, nTop); 1474 } 1475 1476 /* If possible, improve on the iLower estimate using ($P:$L). */ 1477 if( pLower ){ 1478 int n; /* Values extracted from pExpr */ 1479 Expr *pExpr = pLower->pExpr->pRight; 1480 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1481 if( rc==SQLITE_OK && n ){ 1482 tRowcnt iNew; 1483 u16 mask = WO_GT|WO_LE; 1484 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1485 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1486 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1487 if( iNew>iLower ) iLower = iNew; 1488 nOut--; 1489 pLower = 0; 1490 } 1491 } 1492 1493 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1494 if( pUpper ){ 1495 int n; /* Values extracted from pExpr */ 1496 Expr *pExpr = pUpper->pExpr->pRight; 1497 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1498 if( rc==SQLITE_OK && n ){ 1499 tRowcnt iNew; 1500 u16 mask = WO_GT|WO_LE; 1501 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1502 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1503 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1504 if( iNew<iUpper ) iUpper = iNew; 1505 nOut--; 1506 pUpper = 0; 1507 } 1508 } 1509 1510 pBuilder->pRec = pRec; 1511 if( rc==SQLITE_OK ){ 1512 if( iUpper>iLower ){ 1513 nNew = sqlite3LogEst(iUpper - iLower); 1514 /* TUNING: If both iUpper and iLower are derived from the same 1515 ** sample, then assume they are 4x more selective. This brings 1516 ** the estimated selectivity more in line with what it would be 1517 ** if estimated without the use of STAT3/4 tables. */ 1518 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1519 }else{ 1520 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1521 } 1522 if( nNew<nOut ){ 1523 nOut = nNew; 1524 } 1525 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1526 (u32)iLower, (u32)iUpper, nOut)); 1527 } 1528 }else{ 1529 int bDone = 0; 1530 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1531 if( bDone ) return rc; 1532 } 1533 } 1534 #else 1535 UNUSED_PARAMETER(pParse); 1536 UNUSED_PARAMETER(pBuilder); 1537 assert( pLower || pUpper ); 1538 #endif 1539 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1540 nNew = whereRangeAdjust(pLower, nOut); 1541 nNew = whereRangeAdjust(pUpper, nNew); 1542 1543 /* TUNING: If there is both an upper and lower limit and neither limit 1544 ** has an application-defined likelihood(), assume the range is 1545 ** reduced by an additional 75%. This means that, by default, an open-ended 1546 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1547 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1548 ** match 1/64 of the index. */ 1549 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1550 nNew -= 20; 1551 } 1552 1553 nOut -= (pLower!=0) + (pUpper!=0); 1554 if( nNew<10 ) nNew = 10; 1555 if( nNew<nOut ) nOut = nNew; 1556 #if defined(WHERETRACE_ENABLED) 1557 if( pLoop->nOut>nOut ){ 1558 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1559 pLoop->nOut, nOut)); 1560 } 1561 #endif 1562 pLoop->nOut = (LogEst)nOut; 1563 return rc; 1564 } 1565 1566 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1567 /* 1568 ** Estimate the number of rows that will be returned based on 1569 ** an equality constraint x=VALUE and where that VALUE occurs in 1570 ** the histogram data. This only works when x is the left-most 1571 ** column of an index and sqlite_stat3 histogram data is available 1572 ** for that index. When pExpr==NULL that means the constraint is 1573 ** "x IS NULL" instead of "x=VALUE". 1574 ** 1575 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1576 ** If unable to make an estimate, leave *pnRow unchanged and return 1577 ** non-zero. 1578 ** 1579 ** This routine can fail if it is unable to load a collating sequence 1580 ** required for string comparison, or if unable to allocate memory 1581 ** for a UTF conversion required for comparison. The error is stored 1582 ** in the pParse structure. 1583 */ 1584 static int whereEqualScanEst( 1585 Parse *pParse, /* Parsing & code generating context */ 1586 WhereLoopBuilder *pBuilder, 1587 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1588 tRowcnt *pnRow /* Write the revised row estimate here */ 1589 ){ 1590 Index *p = pBuilder->pNew->u.btree.pIndex; 1591 int nEq = pBuilder->pNew->u.btree.nEq; 1592 UnpackedRecord *pRec = pBuilder->pRec; 1593 int rc; /* Subfunction return code */ 1594 tRowcnt a[2]; /* Statistics */ 1595 int bOk; 1596 1597 assert( nEq>=1 ); 1598 assert( nEq<=p->nColumn ); 1599 assert( p->aSample!=0 ); 1600 assert( p->nSample>0 ); 1601 assert( pBuilder->nRecValid<nEq ); 1602 1603 /* If values are not available for all fields of the index to the left 1604 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1605 if( pBuilder->nRecValid<(nEq-1) ){ 1606 return SQLITE_NOTFOUND; 1607 } 1608 1609 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1610 ** below would return the same value. */ 1611 if( nEq>=p->nColumn ){ 1612 *pnRow = 1; 1613 return SQLITE_OK; 1614 } 1615 1616 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1617 pBuilder->pRec = pRec; 1618 if( rc!=SQLITE_OK ) return rc; 1619 if( bOk==0 ) return SQLITE_NOTFOUND; 1620 pBuilder->nRecValid = nEq; 1621 1622 whereKeyStats(pParse, p, pRec, 0, a); 1623 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1624 p->zName, nEq-1, (int)a[1])); 1625 *pnRow = a[1]; 1626 1627 return rc; 1628 } 1629 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1630 1631 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1632 /* 1633 ** Estimate the number of rows that will be returned based on 1634 ** an IN constraint where the right-hand side of the IN operator 1635 ** is a list of values. Example: 1636 ** 1637 ** WHERE x IN (1,2,3,4) 1638 ** 1639 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1640 ** If unable to make an estimate, leave *pnRow unchanged and return 1641 ** non-zero. 1642 ** 1643 ** This routine can fail if it is unable to load a collating sequence 1644 ** required for string comparison, or if unable to allocate memory 1645 ** for a UTF conversion required for comparison. The error is stored 1646 ** in the pParse structure. 1647 */ 1648 static int whereInScanEst( 1649 Parse *pParse, /* Parsing & code generating context */ 1650 WhereLoopBuilder *pBuilder, 1651 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1652 tRowcnt *pnRow /* Write the revised row estimate here */ 1653 ){ 1654 Index *p = pBuilder->pNew->u.btree.pIndex; 1655 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1656 int nRecValid = pBuilder->nRecValid; 1657 int rc = SQLITE_OK; /* Subfunction return code */ 1658 tRowcnt nEst; /* Number of rows for a single term */ 1659 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1660 int i; /* Loop counter */ 1661 1662 assert( p->aSample!=0 ); 1663 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1664 nEst = nRow0; 1665 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1666 nRowEst += nEst; 1667 pBuilder->nRecValid = nRecValid; 1668 } 1669 1670 if( rc==SQLITE_OK ){ 1671 if( nRowEst > nRow0 ) nRowEst = nRow0; 1672 *pnRow = nRowEst; 1673 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1674 } 1675 assert( pBuilder->nRecValid==nRecValid ); 1676 return rc; 1677 } 1678 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1679 1680 1681 #ifdef WHERETRACE_ENABLED 1682 /* 1683 ** Print the content of a WhereTerm object 1684 */ 1685 static void whereTermPrint(WhereTerm *pTerm, int iTerm){ 1686 if( pTerm==0 ){ 1687 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1688 }else{ 1689 char zType[4]; 1690 char zLeft[50]; 1691 memcpy(zType, "...", 4); 1692 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1693 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1694 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1695 if( pTerm->eOperator & WO_SINGLE ){ 1696 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 1697 pTerm->leftCursor, pTerm->u.leftColumn); 1698 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 1699 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%lld", 1700 pTerm->u.pOrInfo->indexable); 1701 }else{ 1702 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 1703 } 1704 sqlite3DebugPrintf( 1705 "TERM-%-3d %p %s %-12s prob=%-3d op=0x%03x wtFlags=0x%04x", 1706 iTerm, pTerm, zType, zLeft, pTerm->truthProb, 1707 pTerm->eOperator, pTerm->wtFlags); 1708 if( pTerm->iField ){ 1709 sqlite3DebugPrintf(" iField=%d\n", pTerm->iField); 1710 }else{ 1711 sqlite3DebugPrintf("\n"); 1712 } 1713 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 1714 } 1715 } 1716 #endif 1717 1718 #ifdef WHERETRACE_ENABLED 1719 /* 1720 ** Show the complete content of a WhereClause 1721 */ 1722 void sqlite3WhereClausePrint(WhereClause *pWC){ 1723 int i; 1724 for(i=0; i<pWC->nTerm; i++){ 1725 whereTermPrint(&pWC->a[i], i); 1726 } 1727 } 1728 #endif 1729 1730 #ifdef WHERETRACE_ENABLED 1731 /* 1732 ** Print a WhereLoop object for debugging purposes 1733 */ 1734 static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){ 1735 WhereInfo *pWInfo = pWC->pWInfo; 1736 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 1737 struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab; 1738 Table *pTab = pItem->pTab; 1739 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 1740 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 1741 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 1742 sqlite3DebugPrintf(" %12s", 1743 pItem->zAlias ? pItem->zAlias : pTab->zName); 1744 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 1745 const char *zName; 1746 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 1747 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 1748 int i = sqlite3Strlen30(zName) - 1; 1749 while( zName[i]!='_' ) i--; 1750 zName += i; 1751 } 1752 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 1753 }else{ 1754 sqlite3DebugPrintf("%20s",""); 1755 } 1756 }else{ 1757 char *z; 1758 if( p->u.vtab.idxStr ){ 1759 z = sqlite3_mprintf("(%d,\"%s\",%x)", 1760 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 1761 }else{ 1762 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 1763 } 1764 sqlite3DebugPrintf(" %-19s", z); 1765 sqlite3_free(z); 1766 } 1767 if( p->wsFlags & WHERE_SKIPSCAN ){ 1768 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 1769 }else{ 1770 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); 1771 } 1772 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 1773 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 1774 int i; 1775 for(i=0; i<p->nLTerm; i++){ 1776 whereTermPrint(p->aLTerm[i], i); 1777 } 1778 } 1779 } 1780 #endif 1781 1782 /* 1783 ** Convert bulk memory into a valid WhereLoop that can be passed 1784 ** to whereLoopClear harmlessly. 1785 */ 1786 static void whereLoopInit(WhereLoop *p){ 1787 p->aLTerm = p->aLTermSpace; 1788 p->nLTerm = 0; 1789 p->nLSlot = ArraySize(p->aLTermSpace); 1790 p->wsFlags = 0; 1791 } 1792 1793 /* 1794 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 1795 */ 1796 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 1797 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 1798 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 1799 sqlite3_free(p->u.vtab.idxStr); 1800 p->u.vtab.needFree = 0; 1801 p->u.vtab.idxStr = 0; 1802 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 1803 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 1804 sqlite3DbFreeNN(db, p->u.btree.pIndex); 1805 p->u.btree.pIndex = 0; 1806 } 1807 } 1808 } 1809 1810 /* 1811 ** Deallocate internal memory used by a WhereLoop object 1812 */ 1813 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 1814 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1815 whereLoopClearUnion(db, p); 1816 whereLoopInit(p); 1817 } 1818 1819 /* 1820 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 1821 */ 1822 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 1823 WhereTerm **paNew; 1824 if( p->nLSlot>=n ) return SQLITE_OK; 1825 n = (n+7)&~7; 1826 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 1827 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 1828 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 1829 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1830 p->aLTerm = paNew; 1831 p->nLSlot = n; 1832 return SQLITE_OK; 1833 } 1834 1835 /* 1836 ** Transfer content from the second pLoop into the first. 1837 */ 1838 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 1839 whereLoopClearUnion(db, pTo); 1840 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 1841 memset(&pTo->u, 0, sizeof(pTo->u)); 1842 return SQLITE_NOMEM_BKPT; 1843 } 1844 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 1845 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 1846 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 1847 pFrom->u.vtab.needFree = 0; 1848 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 1849 pFrom->u.btree.pIndex = 0; 1850 } 1851 return SQLITE_OK; 1852 } 1853 1854 /* 1855 ** Delete a WhereLoop object 1856 */ 1857 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 1858 whereLoopClear(db, p); 1859 sqlite3DbFreeNN(db, p); 1860 } 1861 1862 /* 1863 ** Free a WhereInfo structure 1864 */ 1865 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 1866 if( ALWAYS(pWInfo) ){ 1867 int i; 1868 for(i=0; i<pWInfo->nLevel; i++){ 1869 WhereLevel *pLevel = &pWInfo->a[i]; 1870 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){ 1871 sqlite3DbFree(db, pLevel->u.in.aInLoop); 1872 } 1873 } 1874 sqlite3WhereClauseClear(&pWInfo->sWC); 1875 while( pWInfo->pLoops ){ 1876 WhereLoop *p = pWInfo->pLoops; 1877 pWInfo->pLoops = p->pNextLoop; 1878 whereLoopDelete(db, p); 1879 } 1880 sqlite3DbFreeNN(db, pWInfo); 1881 } 1882 } 1883 1884 /* 1885 ** Return TRUE if all of the following are true: 1886 ** 1887 ** (1) X has the same or lower cost that Y 1888 ** (2) X is a proper subset of Y 1889 ** (3) X skips at least as many columns as Y 1890 ** 1891 ** By "proper subset" we mean that X uses fewer WHERE clause terms 1892 ** than Y and that every WHERE clause term used by X is also used 1893 ** by Y. 1894 ** 1895 ** If X is a proper subset of Y then Y is a better choice and ought 1896 ** to have a lower cost. This routine returns TRUE when that cost 1897 ** relationship is inverted and needs to be adjusted. The third rule 1898 ** was added because if X uses skip-scan less than Y it still might 1899 ** deserve a lower cost even if it is a proper subset of Y. 1900 */ 1901 static int whereLoopCheaperProperSubset( 1902 const WhereLoop *pX, /* First WhereLoop to compare */ 1903 const WhereLoop *pY /* Compare against this WhereLoop */ 1904 ){ 1905 int i, j; 1906 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 1907 return 0; /* X is not a subset of Y */ 1908 } 1909 if( pY->nSkip > pX->nSkip ) return 0; 1910 if( pX->rRun >= pY->rRun ){ 1911 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ 1912 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ 1913 } 1914 for(i=pX->nLTerm-1; i>=0; i--){ 1915 if( pX->aLTerm[i]==0 ) continue; 1916 for(j=pY->nLTerm-1; j>=0; j--){ 1917 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 1918 } 1919 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 1920 } 1921 return 1; /* All conditions meet */ 1922 } 1923 1924 /* 1925 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so 1926 ** that: 1927 ** 1928 ** (1) pTemplate costs less than any other WhereLoops that are a proper 1929 ** subset of pTemplate 1930 ** 1931 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 1932 ** is a proper subset. 1933 ** 1934 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 1935 ** WHERE clause terms than Y and that every WHERE clause term used by X is 1936 ** also used by Y. 1937 */ 1938 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 1939 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 1940 for(; p; p=p->pNextLoop){ 1941 if( p->iTab!=pTemplate->iTab ) continue; 1942 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 1943 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 1944 /* Adjust pTemplate cost downward so that it is cheaper than its 1945 ** subset p. */ 1946 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 1947 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); 1948 pTemplate->rRun = p->rRun; 1949 pTemplate->nOut = p->nOut - 1; 1950 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 1951 /* Adjust pTemplate cost upward so that it is costlier than p since 1952 ** pTemplate is a proper subset of p */ 1953 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 1954 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); 1955 pTemplate->rRun = p->rRun; 1956 pTemplate->nOut = p->nOut + 1; 1957 } 1958 } 1959 } 1960 1961 /* 1962 ** Search the list of WhereLoops in *ppPrev looking for one that can be 1963 ** replaced by pTemplate. 1964 ** 1965 ** Return NULL if pTemplate does not belong on the WhereLoop list. 1966 ** In other words if pTemplate ought to be dropped from further consideration. 1967 ** 1968 ** If pX is a WhereLoop that pTemplate can replace, then return the 1969 ** link that points to pX. 1970 ** 1971 ** If pTemplate cannot replace any existing element of the list but needs 1972 ** to be added to the list as a new entry, then return a pointer to the 1973 ** tail of the list. 1974 */ 1975 static WhereLoop **whereLoopFindLesser( 1976 WhereLoop **ppPrev, 1977 const WhereLoop *pTemplate 1978 ){ 1979 WhereLoop *p; 1980 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 1981 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 1982 /* If either the iTab or iSortIdx values for two WhereLoop are different 1983 ** then those WhereLoops need to be considered separately. Neither is 1984 ** a candidate to replace the other. */ 1985 continue; 1986 } 1987 /* In the current implementation, the rSetup value is either zero 1988 ** or the cost of building an automatic index (NlogN) and the NlogN 1989 ** is the same for compatible WhereLoops. */ 1990 assert( p->rSetup==0 || pTemplate->rSetup==0 1991 || p->rSetup==pTemplate->rSetup ); 1992 1993 /* whereLoopAddBtree() always generates and inserts the automatic index 1994 ** case first. Hence compatible candidate WhereLoops never have a larger 1995 ** rSetup. Call this SETUP-INVARIANT */ 1996 assert( p->rSetup>=pTemplate->rSetup ); 1997 1998 /* Any loop using an appliation-defined index (or PRIMARY KEY or 1999 ** UNIQUE constraint) with one or more == constraints is better 2000 ** than an automatic index. Unless it is a skip-scan. */ 2001 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 2002 && (pTemplate->nSkip)==0 2003 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 2004 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 2005 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2006 ){ 2007 break; 2008 } 2009 2010 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2011 ** discarded. WhereLoop p is better if: 2012 ** (1) p has no more dependencies than pTemplate, and 2013 ** (2) p has an equal or lower cost than pTemplate 2014 */ 2015 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2016 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2017 && p->rRun<=pTemplate->rRun /* (2b) */ 2018 && p->nOut<=pTemplate->nOut /* (2c) */ 2019 ){ 2020 return 0; /* Discard pTemplate */ 2021 } 2022 2023 /* If pTemplate is always better than p, then cause p to be overwritten 2024 ** with pTemplate. pTemplate is better than p if: 2025 ** (1) pTemplate has no more dependences than p, and 2026 ** (2) pTemplate has an equal or lower cost than p. 2027 */ 2028 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2029 && p->rRun>=pTemplate->rRun /* (2a) */ 2030 && p->nOut>=pTemplate->nOut /* (2b) */ 2031 ){ 2032 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2033 break; /* Cause p to be overwritten by pTemplate */ 2034 } 2035 } 2036 return ppPrev; 2037 } 2038 2039 /* 2040 ** Insert or replace a WhereLoop entry using the template supplied. 2041 ** 2042 ** An existing WhereLoop entry might be overwritten if the new template 2043 ** is better and has fewer dependencies. Or the template will be ignored 2044 ** and no insert will occur if an existing WhereLoop is faster and has 2045 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2046 ** added based on the template. 2047 ** 2048 ** If pBuilder->pOrSet is not NULL then we care about only the 2049 ** prerequisites and rRun and nOut costs of the N best loops. That 2050 ** information is gathered in the pBuilder->pOrSet object. This special 2051 ** processing mode is used only for OR clause processing. 2052 ** 2053 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2054 ** still might overwrite similar loops with the new template if the 2055 ** new template is better. Loops may be overwritten if the following 2056 ** conditions are met: 2057 ** 2058 ** (1) They have the same iTab. 2059 ** (2) They have the same iSortIdx. 2060 ** (3) The template has same or fewer dependencies than the current loop 2061 ** (4) The template has the same or lower cost than the current loop 2062 */ 2063 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2064 WhereLoop **ppPrev, *p; 2065 WhereInfo *pWInfo = pBuilder->pWInfo; 2066 sqlite3 *db = pWInfo->pParse->db; 2067 int rc; 2068 2069 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2070 ** and prereqs. 2071 */ 2072 if( pBuilder->pOrSet!=0 ){ 2073 if( pTemplate->nLTerm ){ 2074 #if WHERETRACE_ENABLED 2075 u16 n = pBuilder->pOrSet->n; 2076 int x = 2077 #endif 2078 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2079 pTemplate->nOut); 2080 #if WHERETRACE_ENABLED /* 0x8 */ 2081 if( sqlite3WhereTrace & 0x8 ){ 2082 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2083 whereLoopPrint(pTemplate, pBuilder->pWC); 2084 } 2085 #endif 2086 } 2087 return SQLITE_OK; 2088 } 2089 2090 /* Look for an existing WhereLoop to replace with pTemplate 2091 */ 2092 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2093 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2094 2095 if( ppPrev==0 ){ 2096 /* There already exists a WhereLoop on the list that is better 2097 ** than pTemplate, so just ignore pTemplate */ 2098 #if WHERETRACE_ENABLED /* 0x8 */ 2099 if( sqlite3WhereTrace & 0x8 ){ 2100 sqlite3DebugPrintf(" skip: "); 2101 whereLoopPrint(pTemplate, pBuilder->pWC); 2102 } 2103 #endif 2104 return SQLITE_OK; 2105 }else{ 2106 p = *ppPrev; 2107 } 2108 2109 /* If we reach this point it means that either p[] should be overwritten 2110 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2111 ** WhereLoop and insert it. 2112 */ 2113 #if WHERETRACE_ENABLED /* 0x8 */ 2114 if( sqlite3WhereTrace & 0x8 ){ 2115 if( p!=0 ){ 2116 sqlite3DebugPrintf("replace: "); 2117 whereLoopPrint(p, pBuilder->pWC); 2118 sqlite3DebugPrintf(" with: "); 2119 }else{ 2120 sqlite3DebugPrintf(" add: "); 2121 } 2122 whereLoopPrint(pTemplate, pBuilder->pWC); 2123 } 2124 #endif 2125 if( p==0 ){ 2126 /* Allocate a new WhereLoop to add to the end of the list */ 2127 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2128 if( p==0 ) return SQLITE_NOMEM_BKPT; 2129 whereLoopInit(p); 2130 p->pNextLoop = 0; 2131 }else{ 2132 /* We will be overwriting WhereLoop p[]. But before we do, first 2133 ** go through the rest of the list and delete any other entries besides 2134 ** p[] that are also supplated by pTemplate */ 2135 WhereLoop **ppTail = &p->pNextLoop; 2136 WhereLoop *pToDel; 2137 while( *ppTail ){ 2138 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2139 if( ppTail==0 ) break; 2140 pToDel = *ppTail; 2141 if( pToDel==0 ) break; 2142 *ppTail = pToDel->pNextLoop; 2143 #if WHERETRACE_ENABLED /* 0x8 */ 2144 if( sqlite3WhereTrace & 0x8 ){ 2145 sqlite3DebugPrintf(" delete: "); 2146 whereLoopPrint(pToDel, pBuilder->pWC); 2147 } 2148 #endif 2149 whereLoopDelete(db, pToDel); 2150 } 2151 } 2152 rc = whereLoopXfer(db, p, pTemplate); 2153 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2154 Index *pIndex = p->u.btree.pIndex; 2155 if( pIndex && pIndex->tnum==0 ){ 2156 p->u.btree.pIndex = 0; 2157 } 2158 } 2159 return rc; 2160 } 2161 2162 /* 2163 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2164 ** WHERE clause that reference the loop but which are not used by an 2165 ** index. 2166 * 2167 ** For every WHERE clause term that is not used by the index 2168 ** and which has a truth probability assigned by one of the likelihood(), 2169 ** likely(), or unlikely() SQL functions, reduce the estimated number 2170 ** of output rows by the probability specified. 2171 ** 2172 ** TUNING: For every WHERE clause term that is not used by the index 2173 ** and which does not have an assigned truth probability, heuristics 2174 ** described below are used to try to estimate the truth probability. 2175 ** TODO --> Perhaps this is something that could be improved by better 2176 ** table statistics. 2177 ** 2178 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2179 ** value corresponds to -1 in LogEst notation, so this means decrement 2180 ** the WhereLoop.nOut field for every such WHERE clause term. 2181 ** 2182 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2183 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2184 ** final output row estimate is no greater than 1/4 of the total number 2185 ** of rows in the table. In other words, assume that x==EXPR will filter 2186 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2187 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2188 ** on the "x" column and so in that case only cap the output row estimate 2189 ** at 1/2 instead of 1/4. 2190 */ 2191 static void whereLoopOutputAdjust( 2192 WhereClause *pWC, /* The WHERE clause */ 2193 WhereLoop *pLoop, /* The loop to adjust downward */ 2194 LogEst nRow /* Number of rows in the entire table */ 2195 ){ 2196 WhereTerm *pTerm, *pX; 2197 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2198 int i, j, k; 2199 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2200 2201 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2202 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ 2203 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; 2204 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2205 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2206 for(j=pLoop->nLTerm-1; j>=0; j--){ 2207 pX = pLoop->aLTerm[j]; 2208 if( pX==0 ) continue; 2209 if( pX==pTerm ) break; 2210 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2211 } 2212 if( j<0 ){ 2213 if( pTerm->truthProb<=0 ){ 2214 /* If a truth probability is specified using the likelihood() hints, 2215 ** then use the probability provided by the application. */ 2216 pLoop->nOut += pTerm->truthProb; 2217 }else{ 2218 /* In the absence of explicit truth probabilities, use heuristics to 2219 ** guess a reasonable truth probability. */ 2220 pLoop->nOut--; 2221 if( pTerm->eOperator&(WO_EQ|WO_IS) ){ 2222 Expr *pRight = pTerm->pExpr->pRight; 2223 testcase( pTerm->pExpr->op==TK_IS ); 2224 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2225 k = 10; 2226 }else{ 2227 k = 20; 2228 } 2229 if( iReduce<k ) iReduce = k; 2230 } 2231 } 2232 } 2233 } 2234 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce; 2235 } 2236 2237 /* 2238 ** Term pTerm is a vector range comparison operation. The first comparison 2239 ** in the vector can be optimized using column nEq of the index. This 2240 ** function returns the total number of vector elements that can be used 2241 ** as part of the range comparison. 2242 ** 2243 ** For example, if the query is: 2244 ** 2245 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2246 ** 2247 ** and the index: 2248 ** 2249 ** CREATE INDEX ... ON (a, b, c, d, e) 2250 ** 2251 ** then this function would be invoked with nEq=1. The value returned in 2252 ** this case is 3. 2253 */ 2254 static int whereRangeVectorLen( 2255 Parse *pParse, /* Parsing context */ 2256 int iCur, /* Cursor open on pIdx */ 2257 Index *pIdx, /* The index to be used for a inequality constraint */ 2258 int nEq, /* Number of prior equality constraints on same index */ 2259 WhereTerm *pTerm /* The vector inequality constraint */ 2260 ){ 2261 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2262 int i; 2263 2264 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2265 for(i=1; i<nCmp; i++){ 2266 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2267 ** of the index. If not, exit the loop. */ 2268 char aff; /* Comparison affinity */ 2269 char idxaff = 0; /* Indexed columns affinity */ 2270 CollSeq *pColl; /* Comparison collation sequence */ 2271 Expr *pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2272 Expr *pRhs = pTerm->pExpr->pRight; 2273 if( pRhs->flags & EP_xIsSelect ){ 2274 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2275 }else{ 2276 pRhs = pRhs->x.pList->a[i].pExpr; 2277 } 2278 2279 /* Check that the LHS of the comparison is a column reference to 2280 ** the right column of the right source table. And that the sort 2281 ** order of the index column is the same as the sort order of the 2282 ** leftmost index column. */ 2283 if( pLhs->op!=TK_COLUMN 2284 || pLhs->iTable!=iCur 2285 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2286 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2287 ){ 2288 break; 2289 } 2290 2291 testcase( pLhs->iColumn==XN_ROWID ); 2292 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2293 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2294 if( aff!=idxaff ) break; 2295 2296 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2297 if( pColl==0 ) break; 2298 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2299 } 2300 return i; 2301 } 2302 2303 /* 2304 ** Adjust the cost C by the costMult facter T. This only occurs if 2305 ** compiled with -DSQLITE_ENABLE_COSTMULT 2306 */ 2307 #ifdef SQLITE_ENABLE_COSTMULT 2308 # define ApplyCostMultiplier(C,T) C += T 2309 #else 2310 # define ApplyCostMultiplier(C,T) 2311 #endif 2312 2313 /* 2314 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2315 ** index pIndex. Try to match one more. 2316 ** 2317 ** When this function is called, pBuilder->pNew->nOut contains the 2318 ** number of rows expected to be visited by filtering using the nEq 2319 ** terms only. If it is modified, this value is restored before this 2320 ** function returns. 2321 ** 2322 ** If pProbe->tnum==0, that means pIndex is a fake index used for the 2323 ** INTEGER PRIMARY KEY. 2324 */ 2325 static int whereLoopAddBtreeIndex( 2326 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2327 struct SrcList_item *pSrc, /* FROM clause term being analyzed */ 2328 Index *pProbe, /* An index on pSrc */ 2329 LogEst nInMul /* log(Number of iterations due to IN) */ 2330 ){ 2331 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2332 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2333 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2334 WhereLoop *pNew; /* Template WhereLoop under construction */ 2335 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2336 int opMask; /* Valid operators for constraints */ 2337 WhereScan scan; /* Iterator for WHERE terms */ 2338 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2339 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2340 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2341 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2342 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2343 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2344 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2345 LogEst saved_nOut; /* Original value of pNew->nOut */ 2346 int rc = SQLITE_OK; /* Return code */ 2347 LogEst rSize; /* Number of rows in the table */ 2348 LogEst rLogSize; /* Logarithm of table size */ 2349 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2350 2351 pNew = pBuilder->pNew; 2352 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2353 WHERETRACE(0x800, ("BEGIN addBtreeIdx(%s), nEq=%d\n", 2354 pProbe->zName, pNew->u.btree.nEq)); 2355 2356 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2357 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2358 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2359 opMask = WO_LT|WO_LE; 2360 }else{ 2361 assert( pNew->u.btree.nBtm==0 ); 2362 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2363 } 2364 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2365 2366 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2367 2368 saved_nEq = pNew->u.btree.nEq; 2369 saved_nBtm = pNew->u.btree.nBtm; 2370 saved_nTop = pNew->u.btree.nTop; 2371 saved_nSkip = pNew->nSkip; 2372 saved_nLTerm = pNew->nLTerm; 2373 saved_wsFlags = pNew->wsFlags; 2374 saved_prereq = pNew->prereq; 2375 saved_nOut = pNew->nOut; 2376 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2377 opMask, pProbe); 2378 pNew->rSetup = 0; 2379 rSize = pProbe->aiRowLogEst[0]; 2380 rLogSize = estLog(rSize); 2381 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2382 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2383 LogEst rCostIdx; 2384 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2385 int nIn = 0; 2386 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2387 int nRecValid = pBuilder->nRecValid; 2388 #endif 2389 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2390 && indexColumnNotNull(pProbe, saved_nEq) 2391 ){ 2392 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2393 } 2394 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2395 2396 /* Do not allow the upper bound of a LIKE optimization range constraint 2397 ** to mix with a lower range bound from some other source */ 2398 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2399 2400 /* Do not allow IS constraints from the WHERE clause to be used by the 2401 ** right table of a LEFT JOIN. Only constraints in the ON clause are 2402 ** allowed */ 2403 if( (pSrc->fg.jointype & JT_LEFT)!=0 2404 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2405 && (eOp & (WO_IS|WO_ISNULL))!=0 2406 ){ 2407 testcase( eOp & WO_IS ); 2408 testcase( eOp & WO_ISNULL ); 2409 continue; 2410 } 2411 2412 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2413 pBuilder->bldFlags |= SQLITE_BLDF_UNIQUE; 2414 }else{ 2415 pBuilder->bldFlags |= SQLITE_BLDF_INDEXED; 2416 } 2417 pNew->wsFlags = saved_wsFlags; 2418 pNew->u.btree.nEq = saved_nEq; 2419 pNew->u.btree.nBtm = saved_nBtm; 2420 pNew->u.btree.nTop = saved_nTop; 2421 pNew->nLTerm = saved_nLTerm; 2422 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2423 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2424 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2425 2426 assert( nInMul==0 2427 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2428 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2429 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2430 ); 2431 2432 if( eOp & WO_IN ){ 2433 Expr *pExpr = pTerm->pExpr; 2434 pNew->wsFlags |= WHERE_COLUMN_IN; 2435 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 2436 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2437 int i; 2438 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2439 2440 /* The expression may actually be of the form (x, y) IN (SELECT...). 2441 ** In this case there is a separate term for each of (x) and (y). 2442 ** However, the nIn multiplier should only be applied once, not once 2443 ** for each such term. The following loop checks that pTerm is the 2444 ** first such term in use, and sets nIn back to 0 if it is not. */ 2445 for(i=0; i<pNew->nLTerm-1; i++){ 2446 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2447 } 2448 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2449 /* "x IN (value, value, ...)" */ 2450 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2451 assert( nIn>0 ); /* RHS always has 2 or more terms... The parser 2452 ** changes "x IN (?)" into "x=?". */ 2453 } 2454 }else if( eOp & (WO_EQ|WO_IS) ){ 2455 int iCol = pProbe->aiColumn[saved_nEq]; 2456 pNew->wsFlags |= WHERE_COLUMN_EQ; 2457 assert( saved_nEq==pNew->u.btree.nEq ); 2458 if( iCol==XN_ROWID 2459 || (iCol>0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2460 ){ 2461 if( iCol>=0 && pProbe->uniqNotNull==0 ){ 2462 pNew->wsFlags |= WHERE_UNQ_WANTED; 2463 }else{ 2464 pNew->wsFlags |= WHERE_ONEROW; 2465 } 2466 } 2467 }else if( eOp & WO_ISNULL ){ 2468 pNew->wsFlags |= WHERE_COLUMN_NULL; 2469 }else if( eOp & (WO_GT|WO_GE) ){ 2470 testcase( eOp & WO_GT ); 2471 testcase( eOp & WO_GE ); 2472 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2473 pNew->u.btree.nBtm = whereRangeVectorLen( 2474 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2475 ); 2476 pBtm = pTerm; 2477 pTop = 0; 2478 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2479 /* Range contraints that come from the LIKE optimization are 2480 ** always used in pairs. */ 2481 pTop = &pTerm[1]; 2482 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2483 assert( pTop->wtFlags & TERM_LIKEOPT ); 2484 assert( pTop->eOperator==WO_LT ); 2485 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2486 pNew->aLTerm[pNew->nLTerm++] = pTop; 2487 pNew->wsFlags |= WHERE_TOP_LIMIT; 2488 pNew->u.btree.nTop = 1; 2489 } 2490 }else{ 2491 assert( eOp & (WO_LT|WO_LE) ); 2492 testcase( eOp & WO_LT ); 2493 testcase( eOp & WO_LE ); 2494 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2495 pNew->u.btree.nTop = whereRangeVectorLen( 2496 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2497 ); 2498 pTop = pTerm; 2499 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2500 pNew->aLTerm[pNew->nLTerm-2] : 0; 2501 } 2502 2503 /* At this point pNew->nOut is set to the number of rows expected to 2504 ** be visited by the index scan before considering term pTerm, or the 2505 ** values of nIn and nInMul. In other words, assuming that all 2506 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2507 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2508 assert( pNew->nOut==saved_nOut ); 2509 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2510 /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4 2511 ** data, using some other estimate. */ 2512 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2513 }else{ 2514 int nEq = ++pNew->u.btree.nEq; 2515 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2516 2517 assert( pNew->nOut==saved_nOut ); 2518 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2519 assert( (eOp & WO_IN) || nIn==0 ); 2520 testcase( eOp & WO_IN ); 2521 pNew->nOut += pTerm->truthProb; 2522 pNew->nOut -= nIn; 2523 }else{ 2524 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2525 tRowcnt nOut = 0; 2526 if( nInMul==0 2527 && pProbe->nSample 2528 && pNew->u.btree.nEq<=pProbe->nSampleCol 2529 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) 2530 ){ 2531 Expr *pExpr = pTerm->pExpr; 2532 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2533 testcase( eOp & WO_EQ ); 2534 testcase( eOp & WO_IS ); 2535 testcase( eOp & WO_ISNULL ); 2536 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2537 }else{ 2538 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2539 } 2540 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2541 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2542 if( nOut ){ 2543 pNew->nOut = sqlite3LogEst(nOut); 2544 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2545 pNew->nOut -= nIn; 2546 } 2547 } 2548 if( nOut==0 ) 2549 #endif 2550 { 2551 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2552 if( eOp & WO_ISNULL ){ 2553 /* TUNING: If there is no likelihood() value, assume that a 2554 ** "col IS NULL" expression matches twice as many rows 2555 ** as (col=?). */ 2556 pNew->nOut += 10; 2557 } 2558 } 2559 } 2560 } 2561 2562 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2563 ** it to pNew->rRun, which is currently set to the cost of the index 2564 ** seek only. Then, if this is a non-covering index, add the cost of 2565 ** visiting the rows in the main table. */ 2566 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2567 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2568 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2569 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2570 } 2571 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2572 2573 nOutUnadjusted = pNew->nOut; 2574 pNew->rRun += nInMul + nIn; 2575 pNew->nOut += nInMul + nIn; 2576 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 2577 rc = whereLoopInsert(pBuilder, pNew); 2578 2579 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2580 pNew->nOut = saved_nOut; 2581 }else{ 2582 pNew->nOut = nOutUnadjusted; 2583 } 2584 2585 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 2586 && pNew->u.btree.nEq<pProbe->nColumn 2587 ){ 2588 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 2589 } 2590 pNew->nOut = saved_nOut; 2591 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2592 pBuilder->nRecValid = nRecValid; 2593 #endif 2594 } 2595 pNew->prereq = saved_prereq; 2596 pNew->u.btree.nEq = saved_nEq; 2597 pNew->u.btree.nBtm = saved_nBtm; 2598 pNew->u.btree.nTop = saved_nTop; 2599 pNew->nSkip = saved_nSkip; 2600 pNew->wsFlags = saved_wsFlags; 2601 pNew->nOut = saved_nOut; 2602 pNew->nLTerm = saved_nLTerm; 2603 2604 /* Consider using a skip-scan if there are no WHERE clause constraints 2605 ** available for the left-most terms of the index, and if the average 2606 ** number of repeats in the left-most terms is at least 18. 2607 ** 2608 ** The magic number 18 is selected on the basis that scanning 17 rows 2609 ** is almost always quicker than an index seek (even though if the index 2610 ** contains fewer than 2^17 rows we assume otherwise in other parts of 2611 ** the code). And, even if it is not, it should not be too much slower. 2612 ** On the other hand, the extra seeks could end up being significantly 2613 ** more expensive. */ 2614 assert( 42==sqlite3LogEst(18) ); 2615 if( saved_nEq==saved_nSkip 2616 && saved_nEq+1<pProbe->nKeyCol 2617 && pProbe->noSkipScan==0 2618 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 2619 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 2620 ){ 2621 LogEst nIter; 2622 pNew->u.btree.nEq++; 2623 pNew->nSkip++; 2624 pNew->aLTerm[pNew->nLTerm++] = 0; 2625 pNew->wsFlags |= WHERE_SKIPSCAN; 2626 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 2627 pNew->nOut -= nIter; 2628 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 2629 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 2630 nIter += 5; 2631 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 2632 pNew->nOut = saved_nOut; 2633 pNew->u.btree.nEq = saved_nEq; 2634 pNew->nSkip = saved_nSkip; 2635 pNew->wsFlags = saved_wsFlags; 2636 } 2637 2638 WHERETRACE(0x800, ("END addBtreeIdx(%s), nEq=%d, rc=%d\n", 2639 pProbe->zName, saved_nEq, rc)); 2640 return rc; 2641 } 2642 2643 /* 2644 ** Return True if it is possible that pIndex might be useful in 2645 ** implementing the ORDER BY clause in pBuilder. 2646 ** 2647 ** Return False if pBuilder does not contain an ORDER BY clause or 2648 ** if there is no way for pIndex to be useful in implementing that 2649 ** ORDER BY clause. 2650 */ 2651 static int indexMightHelpWithOrderBy( 2652 WhereLoopBuilder *pBuilder, 2653 Index *pIndex, 2654 int iCursor 2655 ){ 2656 ExprList *pOB; 2657 ExprList *aColExpr; 2658 int ii, jj; 2659 2660 if( pIndex->bUnordered ) return 0; 2661 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 2662 for(ii=0; ii<pOB->nExpr; ii++){ 2663 Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr); 2664 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 2665 if( pExpr->iColumn<0 ) return 1; 2666 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2667 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 2668 } 2669 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 2670 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2671 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 2672 if( sqlite3ExprCompare(0, pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 2673 return 1; 2674 } 2675 } 2676 } 2677 } 2678 return 0; 2679 } 2680 2681 /* 2682 ** Return a bitmask where 1s indicate that the corresponding column of 2683 ** the table is used by an index. Only the first 63 columns are considered. 2684 */ 2685 static Bitmask columnsInIndex(Index *pIdx){ 2686 Bitmask m = 0; 2687 int j; 2688 for(j=pIdx->nColumn-1; j>=0; j--){ 2689 int x = pIdx->aiColumn[j]; 2690 if( x>=0 ){ 2691 testcase( x==BMS-1 ); 2692 testcase( x==BMS-2 ); 2693 if( x<BMS-1 ) m |= MASKBIT(x); 2694 } 2695 } 2696 return m; 2697 } 2698 2699 /* Check to see if a partial index with pPartIndexWhere can be used 2700 ** in the current query. Return true if it can be and false if not. 2701 */ 2702 static int whereUsablePartialIndex(int iTab, WhereClause *pWC, Expr *pWhere){ 2703 int i; 2704 WhereTerm *pTerm; 2705 Parse *pParse = pWC->pWInfo->pParse; 2706 while( pWhere->op==TK_AND ){ 2707 if( !whereUsablePartialIndex(iTab,pWC,pWhere->pLeft) ) return 0; 2708 pWhere = pWhere->pRight; 2709 } 2710 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 2711 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2712 Expr *pExpr = pTerm->pExpr; 2713 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 2714 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 2715 ){ 2716 return 1; 2717 } 2718 } 2719 return 0; 2720 } 2721 2722 /* 2723 ** Add all WhereLoop objects for a single table of the join where the table 2724 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 2725 ** a b-tree table, not a virtual table. 2726 ** 2727 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 2728 ** are calculated as follows: 2729 ** 2730 ** For a full scan, assuming the table (or index) contains nRow rows: 2731 ** 2732 ** cost = nRow * 3.0 // full-table scan 2733 ** cost = nRow * K // scan of covering index 2734 ** cost = nRow * (K+3.0) // scan of non-covering index 2735 ** 2736 ** where K is a value between 1.1 and 3.0 set based on the relative 2737 ** estimated average size of the index and table records. 2738 ** 2739 ** For an index scan, where nVisit is the number of index rows visited 2740 ** by the scan, and nSeek is the number of seek operations required on 2741 ** the index b-tree: 2742 ** 2743 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 2744 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 2745 ** 2746 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 2747 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 2748 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 2749 ** 2750 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 2751 ** of uncertainty. For this reason, scoring is designed to pick plans that 2752 ** "do the least harm" if the estimates are inaccurate. For example, a 2753 ** log(nRow) factor is omitted from a non-covering index scan in order to 2754 ** bias the scoring in favor of using an index, since the worst-case 2755 ** performance of using an index is far better than the worst-case performance 2756 ** of a full table scan. 2757 */ 2758 static int whereLoopAddBtree( 2759 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2760 Bitmask mPrereq /* Extra prerequesites for using this table */ 2761 ){ 2762 WhereInfo *pWInfo; /* WHERE analysis context */ 2763 Index *pProbe; /* An index we are evaluating */ 2764 Index sPk; /* A fake index object for the primary key */ 2765 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 2766 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 2767 SrcList *pTabList; /* The FROM clause */ 2768 struct SrcList_item *pSrc; /* The FROM clause btree term to add */ 2769 WhereLoop *pNew; /* Template WhereLoop object */ 2770 int rc = SQLITE_OK; /* Return code */ 2771 int iSortIdx = 1; /* Index number */ 2772 int b; /* A boolean value */ 2773 LogEst rSize; /* number of rows in the table */ 2774 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 2775 WhereClause *pWC; /* The parsed WHERE clause */ 2776 Table *pTab; /* Table being queried */ 2777 2778 pNew = pBuilder->pNew; 2779 pWInfo = pBuilder->pWInfo; 2780 pTabList = pWInfo->pTabList; 2781 pSrc = pTabList->a + pNew->iTab; 2782 pTab = pSrc->pTab; 2783 pWC = pBuilder->pWC; 2784 assert( !IsVirtual(pSrc->pTab) ); 2785 2786 if( pSrc->pIBIndex ){ 2787 /* An INDEXED BY clause specifies a particular index to use */ 2788 pProbe = pSrc->pIBIndex; 2789 }else if( !HasRowid(pTab) ){ 2790 pProbe = pTab->pIndex; 2791 }else{ 2792 /* There is no INDEXED BY clause. Create a fake Index object in local 2793 ** variable sPk to represent the rowid primary key index. Make this 2794 ** fake index the first in a chain of Index objects with all of the real 2795 ** indices to follow */ 2796 Index *pFirst; /* First of real indices on the table */ 2797 memset(&sPk, 0, sizeof(Index)); 2798 sPk.nKeyCol = 1; 2799 sPk.nColumn = 1; 2800 sPk.aiColumn = &aiColumnPk; 2801 sPk.aiRowLogEst = aiRowEstPk; 2802 sPk.onError = OE_Replace; 2803 sPk.pTable = pTab; 2804 sPk.szIdxRow = pTab->szTabRow; 2805 aiRowEstPk[0] = pTab->nRowLogEst; 2806 aiRowEstPk[1] = 0; 2807 pFirst = pSrc->pTab->pIndex; 2808 if( pSrc->fg.notIndexed==0 ){ 2809 /* The real indices of the table are only considered if the 2810 ** NOT INDEXED qualifier is omitted from the FROM clause */ 2811 sPk.pNext = pFirst; 2812 } 2813 pProbe = &sPk; 2814 } 2815 rSize = pTab->nRowLogEst; 2816 rLogSize = estLog(rSize); 2817 2818 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 2819 /* Automatic indexes */ 2820 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 2821 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 2822 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 2823 && pSrc->pIBIndex==0 /* Has no INDEXED BY clause */ 2824 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 2825 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 2826 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 2827 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 2828 ){ 2829 /* Generate auto-index WhereLoops */ 2830 WhereTerm *pTerm; 2831 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 2832 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 2833 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2834 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 2835 pNew->u.btree.nEq = 1; 2836 pNew->nSkip = 0; 2837 pNew->u.btree.pIndex = 0; 2838 pNew->nLTerm = 1; 2839 pNew->aLTerm[0] = pTerm; 2840 /* TUNING: One-time cost for computing the automatic index is 2841 ** estimated to be X*N*log2(N) where N is the number of rows in 2842 ** the table being indexed and where X is 7 (LogEst=28) for normal 2843 ** tables or 1.375 (LogEst=4) for views and subqueries. The value 2844 ** of X is smaller for views and subqueries so that the query planner 2845 ** will be more aggressive about generating automatic indexes for 2846 ** those objects, since there is no opportunity to add schema 2847 ** indexes on subqueries and views. */ 2848 pNew->rSetup = rLogSize + rSize + 4; 2849 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ 2850 pNew->rSetup += 24; 2851 } 2852 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 2853 if( pNew->rSetup<0 ) pNew->rSetup = 0; 2854 /* TUNING: Each index lookup yields 20 rows in the table. This 2855 ** is more than the usual guess of 10 rows, since we have no way 2856 ** of knowing how selective the index will ultimately be. It would 2857 ** not be unreasonable to make this value much larger. */ 2858 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 2859 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 2860 pNew->wsFlags = WHERE_AUTO_INDEX; 2861 pNew->prereq = mPrereq | pTerm->prereqRight; 2862 rc = whereLoopInsert(pBuilder, pNew); 2863 } 2864 } 2865 } 2866 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 2867 2868 /* Loop over all indices 2869 */ 2870 for(; rc==SQLITE_OK && pProbe; pProbe=pProbe->pNext, iSortIdx++){ 2871 if( pProbe->pPartIdxWhere!=0 2872 && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){ 2873 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 2874 continue; /* Partial index inappropriate for this query */ 2875 } 2876 rSize = pProbe->aiRowLogEst[0]; 2877 pNew->u.btree.nEq = 0; 2878 pNew->u.btree.nBtm = 0; 2879 pNew->u.btree.nTop = 0; 2880 pNew->nSkip = 0; 2881 pNew->nLTerm = 0; 2882 pNew->iSortIdx = 0; 2883 pNew->rSetup = 0; 2884 pNew->prereq = mPrereq; 2885 pNew->nOut = rSize; 2886 pNew->u.btree.pIndex = pProbe; 2887 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 2888 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 2889 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 2890 if( pProbe->tnum<=0 ){ 2891 /* Integer primary key index */ 2892 pNew->wsFlags = WHERE_IPK; 2893 2894 /* Full table scan */ 2895 pNew->iSortIdx = b ? iSortIdx : 0; 2896 /* TUNING: Cost of full table scan is (N*3.0). */ 2897 pNew->rRun = rSize + 16; 2898 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2899 whereLoopOutputAdjust(pWC, pNew, rSize); 2900 rc = whereLoopInsert(pBuilder, pNew); 2901 pNew->nOut = rSize; 2902 if( rc ) break; 2903 }else{ 2904 Bitmask m; 2905 if( pProbe->isCovering ){ 2906 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 2907 m = 0; 2908 }else{ 2909 m = pSrc->colUsed & ~columnsInIndex(pProbe); 2910 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 2911 } 2912 2913 /* Full scan via index */ 2914 if( b 2915 || !HasRowid(pTab) 2916 || pProbe->pPartIdxWhere!=0 2917 || ( m==0 2918 && pProbe->bUnordered==0 2919 && (pProbe->szIdxRow<pTab->szTabRow) 2920 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 2921 && sqlite3GlobalConfig.bUseCis 2922 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 2923 ) 2924 ){ 2925 pNew->iSortIdx = b ? iSortIdx : 0; 2926 2927 /* The cost of visiting the index rows is N*K, where K is 2928 ** between 1.1 and 3.0, depending on the relative sizes of the 2929 ** index and table rows. */ 2930 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 2931 if( m!=0 ){ 2932 /* If this is a non-covering index scan, add in the cost of 2933 ** doing table lookups. The cost will be 3x the number of 2934 ** lookups. Take into account WHERE clause terms that can be 2935 ** satisfied using just the index, and that do not require a 2936 ** table lookup. */ 2937 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 2938 int ii; 2939 int iCur = pSrc->iCursor; 2940 WhereClause *pWC2 = &pWInfo->sWC; 2941 for(ii=0; ii<pWC2->nTerm; ii++){ 2942 WhereTerm *pTerm = &pWC2->a[ii]; 2943 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 2944 break; 2945 } 2946 /* pTerm can be evaluated using just the index. So reduce 2947 ** the expected number of table lookups accordingly */ 2948 if( pTerm->truthProb<=0 ){ 2949 nLookup += pTerm->truthProb; 2950 }else{ 2951 nLookup--; 2952 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 2953 } 2954 } 2955 2956 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 2957 } 2958 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2959 whereLoopOutputAdjust(pWC, pNew, rSize); 2960 rc = whereLoopInsert(pBuilder, pNew); 2961 pNew->nOut = rSize; 2962 if( rc ) break; 2963 } 2964 } 2965 2966 pBuilder->bldFlags = 0; 2967 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 2968 if( pBuilder->bldFlags==SQLITE_BLDF_INDEXED ){ 2969 /* If a non-unique index is used, or if a prefix of the key for 2970 ** unique index is used (making the index functionally non-unique) 2971 ** then the sqlite_stat1 data becomes important for scoring the 2972 ** plan */ 2973 pTab->tabFlags |= TF_StatsUsed; 2974 } 2975 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2976 sqlite3Stat4ProbeFree(pBuilder->pRec); 2977 pBuilder->nRecValid = 0; 2978 pBuilder->pRec = 0; 2979 #endif 2980 2981 /* If there was an INDEXED BY clause, then only that one index is 2982 ** considered. */ 2983 if( pSrc->pIBIndex ) break; 2984 } 2985 return rc; 2986 } 2987 2988 #ifndef SQLITE_OMIT_VIRTUALTABLE 2989 2990 /* 2991 ** Argument pIdxInfo is already populated with all constraints that may 2992 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 2993 ** function marks a subset of those constraints usable, invokes the 2994 ** xBestIndex method and adds the returned plan to pBuilder. 2995 ** 2996 ** A constraint is marked usable if: 2997 ** 2998 ** * Argument mUsable indicates that its prerequisites are available, and 2999 ** 3000 ** * It is not one of the operators specified in the mExclude mask passed 3001 ** as the fourth argument (which in practice is either WO_IN or 0). 3002 ** 3003 ** Argument mPrereq is a mask of tables that must be scanned before the 3004 ** virtual table in question. These are added to the plans prerequisites 3005 ** before it is added to pBuilder. 3006 ** 3007 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3008 ** uses one or more WO_IN terms, or false otherwise. 3009 */ 3010 static int whereLoopAddVirtualOne( 3011 WhereLoopBuilder *pBuilder, 3012 Bitmask mPrereq, /* Mask of tables that must be used. */ 3013 Bitmask mUsable, /* Mask of usable tables */ 3014 u16 mExclude, /* Exclude terms using these operators */ 3015 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3016 u16 mNoOmit, /* Do not omit these constraints */ 3017 int *pbIn /* OUT: True if plan uses an IN(...) op */ 3018 ){ 3019 WhereClause *pWC = pBuilder->pWC; 3020 struct sqlite3_index_constraint *pIdxCons; 3021 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3022 int i; 3023 int mxTerm; 3024 int rc = SQLITE_OK; 3025 WhereLoop *pNew = pBuilder->pNew; 3026 Parse *pParse = pBuilder->pWInfo->pParse; 3027 struct SrcList_item *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3028 int nConstraint = pIdxInfo->nConstraint; 3029 3030 assert( (mUsable & mPrereq)==mPrereq ); 3031 *pbIn = 0; 3032 pNew->prereq = mPrereq; 3033 3034 /* Set the usable flag on the subset of constraints identified by 3035 ** arguments mUsable and mExclude. */ 3036 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3037 for(i=0; i<nConstraint; i++, pIdxCons++){ 3038 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3039 pIdxCons->usable = 0; 3040 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3041 && (pTerm->eOperator & mExclude)==0 3042 ){ 3043 pIdxCons->usable = 1; 3044 } 3045 } 3046 3047 /* Initialize the output fields of the sqlite3_index_info structure */ 3048 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3049 assert( pIdxInfo->needToFreeIdxStr==0 ); 3050 pIdxInfo->idxStr = 0; 3051 pIdxInfo->idxNum = 0; 3052 pIdxInfo->orderByConsumed = 0; 3053 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3054 pIdxInfo->estimatedRows = 25; 3055 pIdxInfo->idxFlags = 0; 3056 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3057 3058 /* Invoke the virtual table xBestIndex() method */ 3059 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3060 if( rc ) return rc; 3061 3062 mxTerm = -1; 3063 assert( pNew->nLSlot>=nConstraint ); 3064 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 3065 pNew->u.vtab.omitMask = 0; 3066 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3067 for(i=0; i<nConstraint; i++, pIdxCons++){ 3068 int iTerm; 3069 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3070 WhereTerm *pTerm; 3071 int j = pIdxCons->iTermOffset; 3072 if( iTerm>=nConstraint 3073 || j<0 3074 || j>=pWC->nTerm 3075 || pNew->aLTerm[iTerm]!=0 3076 || pIdxCons->usable==0 3077 ){ 3078 rc = SQLITE_ERROR; 3079 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3080 return rc; 3081 } 3082 testcase( iTerm==nConstraint-1 ); 3083 testcase( j==0 ); 3084 testcase( j==pWC->nTerm-1 ); 3085 pTerm = &pWC->a[j]; 3086 pNew->prereq |= pTerm->prereqRight; 3087 assert( iTerm<pNew->nLSlot ); 3088 pNew->aLTerm[iTerm] = pTerm; 3089 if( iTerm>mxTerm ) mxTerm = iTerm; 3090 testcase( iTerm==15 ); 3091 testcase( iTerm==16 ); 3092 if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<<iTerm; 3093 if( (pTerm->eOperator & WO_IN)!=0 ){ 3094 /* A virtual table that is constrained by an IN clause may not 3095 ** consume the ORDER BY clause because (1) the order of IN terms 3096 ** is not necessarily related to the order of output terms and 3097 ** (2) Multiple outputs from a single IN value will not merge 3098 ** together. */ 3099 pIdxInfo->orderByConsumed = 0; 3100 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3101 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3102 } 3103 } 3104 } 3105 pNew->u.vtab.omitMask &= ~mNoOmit; 3106 3107 pNew->nLTerm = mxTerm+1; 3108 assert( pNew->nLTerm<=pNew->nLSlot ); 3109 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3110 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3111 pIdxInfo->needToFreeIdxStr = 0; 3112 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3113 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3114 pIdxInfo->nOrderBy : 0); 3115 pNew->rSetup = 0; 3116 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3117 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3118 3119 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3120 ** that the scan will visit at most one row. Clear it otherwise. */ 3121 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3122 pNew->wsFlags |= WHERE_ONEROW; 3123 }else{ 3124 pNew->wsFlags &= ~WHERE_ONEROW; 3125 } 3126 rc = whereLoopInsert(pBuilder, pNew); 3127 if( pNew->u.vtab.needFree ){ 3128 sqlite3_free(pNew->u.vtab.idxStr); 3129 pNew->u.vtab.needFree = 0; 3130 } 3131 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3132 *pbIn, (sqlite3_uint64)mPrereq, 3133 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3134 3135 return rc; 3136 } 3137 3138 3139 /* 3140 ** Add all WhereLoop objects for a table of the join identified by 3141 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3142 ** 3143 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3144 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3145 ** entries that occur before the virtual table in the FROM clause and are 3146 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3147 ** mUnusable mask contains all FROM clause entries that occur after the 3148 ** virtual table and are separated from it by at least one LEFT or 3149 ** CROSS JOIN. 3150 ** 3151 ** For example, if the query were: 3152 ** 3153 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3154 ** 3155 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3156 ** 3157 ** All the tables in mPrereq must be scanned before the current virtual 3158 ** table. So any terms for which all prerequisites are satisfied by 3159 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3160 ** Conversely, all tables in mUnusable must be scanned after the current 3161 ** virtual table, so any terms for which the prerequisites overlap with 3162 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3163 */ 3164 static int whereLoopAddVirtual( 3165 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3166 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3167 Bitmask mUnusable /* Tables that must be scanned after this one */ 3168 ){ 3169 int rc = SQLITE_OK; /* Return code */ 3170 WhereInfo *pWInfo; /* WHERE analysis context */ 3171 Parse *pParse; /* The parsing context */ 3172 WhereClause *pWC; /* The WHERE clause */ 3173 struct SrcList_item *pSrc; /* The FROM clause term to search */ 3174 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3175 int nConstraint; /* Number of constraints in p */ 3176 int bIn; /* True if plan uses IN(...) operator */ 3177 WhereLoop *pNew; 3178 Bitmask mBest; /* Tables used by best possible plan */ 3179 u16 mNoOmit; 3180 3181 assert( (mPrereq & mUnusable)==0 ); 3182 pWInfo = pBuilder->pWInfo; 3183 pParse = pWInfo->pParse; 3184 pWC = pBuilder->pWC; 3185 pNew = pBuilder->pNew; 3186 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3187 assert( IsVirtual(pSrc->pTab) ); 3188 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy, 3189 &mNoOmit); 3190 if( p==0 ) return SQLITE_NOMEM_BKPT; 3191 pNew->rSetup = 0; 3192 pNew->wsFlags = WHERE_VIRTUALTABLE; 3193 pNew->nLTerm = 0; 3194 pNew->u.vtab.needFree = 0; 3195 nConstraint = p->nConstraint; 3196 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3197 sqlite3DbFree(pParse->db, p); 3198 return SQLITE_NOMEM_BKPT; 3199 } 3200 3201 /* First call xBestIndex() with all constraints usable. */ 3202 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3203 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn); 3204 3205 /* If the call to xBestIndex() with all terms enabled produced a plan 3206 ** that does not require any source tables (IOW: a plan with mBest==0), 3207 ** then there is no point in making any further calls to xBestIndex() 3208 ** since they will all return the same result (if the xBestIndex() 3209 ** implementation is sane). */ 3210 if( rc==SQLITE_OK && (mBest = (pNew->prereq & ~mPrereq))!=0 ){ 3211 int seenZero = 0; /* True if a plan with no prereqs seen */ 3212 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3213 Bitmask mPrev = 0; 3214 Bitmask mBestNoIn = 0; 3215 3216 /* If the plan produced by the earlier call uses an IN(...) term, call 3217 ** xBestIndex again, this time with IN(...) terms disabled. */ 3218 if( bIn ){ 3219 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3220 rc = whereLoopAddVirtualOne( 3221 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn); 3222 assert( bIn==0 ); 3223 mBestNoIn = pNew->prereq & ~mPrereq; 3224 if( mBestNoIn==0 ){ 3225 seenZero = 1; 3226 seenZeroNoIN = 1; 3227 } 3228 } 3229 3230 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3231 ** in the set of terms that apply to the current virtual table. */ 3232 while( rc==SQLITE_OK ){ 3233 int i; 3234 Bitmask mNext = ALLBITS; 3235 assert( mNext>0 ); 3236 for(i=0; i<nConstraint; i++){ 3237 Bitmask mThis = ( 3238 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3239 ); 3240 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3241 } 3242 mPrev = mNext; 3243 if( mNext==ALLBITS ) break; 3244 if( mNext==mBest || mNext==mBestNoIn ) continue; 3245 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3246 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3247 rc = whereLoopAddVirtualOne( 3248 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn); 3249 if( pNew->prereq==mPrereq ){ 3250 seenZero = 1; 3251 if( bIn==0 ) seenZeroNoIN = 1; 3252 } 3253 } 3254 3255 /* If the calls to xBestIndex() in the above loop did not find a plan 3256 ** that requires no source tables at all (i.e. one guaranteed to be 3257 ** usable), make a call here with all source tables disabled */ 3258 if( rc==SQLITE_OK && seenZero==0 ){ 3259 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3260 rc = whereLoopAddVirtualOne( 3261 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn); 3262 if( bIn==0 ) seenZeroNoIN = 1; 3263 } 3264 3265 /* If the calls to xBestIndex() have so far failed to find a plan 3266 ** that requires no source tables at all and does not use an IN(...) 3267 ** operator, make a final call to obtain one here. */ 3268 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3269 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3270 rc = whereLoopAddVirtualOne( 3271 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn); 3272 } 3273 } 3274 3275 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3276 sqlite3DbFreeNN(pParse->db, p); 3277 return rc; 3278 } 3279 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3280 3281 /* 3282 ** Add WhereLoop entries to handle OR terms. This works for either 3283 ** btrees or virtual tables. 3284 */ 3285 static int whereLoopAddOr( 3286 WhereLoopBuilder *pBuilder, 3287 Bitmask mPrereq, 3288 Bitmask mUnusable 3289 ){ 3290 WhereInfo *pWInfo = pBuilder->pWInfo; 3291 WhereClause *pWC; 3292 WhereLoop *pNew; 3293 WhereTerm *pTerm, *pWCEnd; 3294 int rc = SQLITE_OK; 3295 int iCur; 3296 WhereClause tempWC; 3297 WhereLoopBuilder sSubBuild; 3298 WhereOrSet sSum, sCur; 3299 struct SrcList_item *pItem; 3300 3301 pWC = pBuilder->pWC; 3302 pWCEnd = pWC->a + pWC->nTerm; 3303 pNew = pBuilder->pNew; 3304 memset(&sSum, 0, sizeof(sSum)); 3305 pItem = pWInfo->pTabList->a + pNew->iTab; 3306 iCur = pItem->iCursor; 3307 3308 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3309 if( (pTerm->eOperator & WO_OR)!=0 3310 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3311 ){ 3312 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3313 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3314 WhereTerm *pOrTerm; 3315 int once = 1; 3316 int i, j; 3317 3318 sSubBuild = *pBuilder; 3319 sSubBuild.pOrderBy = 0; 3320 sSubBuild.pOrSet = &sCur; 3321 3322 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3323 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3324 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3325 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3326 }else if( pOrTerm->leftCursor==iCur ){ 3327 tempWC.pWInfo = pWC->pWInfo; 3328 tempWC.pOuter = pWC; 3329 tempWC.op = TK_AND; 3330 tempWC.nTerm = 1; 3331 tempWC.a = pOrTerm; 3332 sSubBuild.pWC = &tempWC; 3333 }else{ 3334 continue; 3335 } 3336 sCur.n = 0; 3337 #ifdef WHERETRACE_ENABLED 3338 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3339 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3340 if( sqlite3WhereTrace & 0x400 ){ 3341 sqlite3WhereClausePrint(sSubBuild.pWC); 3342 } 3343 #endif 3344 #ifndef SQLITE_OMIT_VIRTUALTABLE 3345 if( IsVirtual(pItem->pTab) ){ 3346 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 3347 }else 3348 #endif 3349 { 3350 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 3351 } 3352 if( rc==SQLITE_OK ){ 3353 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 3354 } 3355 assert( rc==SQLITE_OK || sCur.n==0 ); 3356 if( sCur.n==0 ){ 3357 sSum.n = 0; 3358 break; 3359 }else if( once ){ 3360 whereOrMove(&sSum, &sCur); 3361 once = 0; 3362 }else{ 3363 WhereOrSet sPrev; 3364 whereOrMove(&sPrev, &sSum); 3365 sSum.n = 0; 3366 for(i=0; i<sPrev.n; i++){ 3367 for(j=0; j<sCur.n; j++){ 3368 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3369 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3370 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3371 } 3372 } 3373 } 3374 } 3375 pNew->nLTerm = 1; 3376 pNew->aLTerm[0] = pTerm; 3377 pNew->wsFlags = WHERE_MULTI_OR; 3378 pNew->rSetup = 0; 3379 pNew->iSortIdx = 0; 3380 memset(&pNew->u, 0, sizeof(pNew->u)); 3381 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3382 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3383 ** of all sub-scans required by the OR-scan. However, due to rounding 3384 ** errors, it may be that the cost of the OR-scan is equal to its 3385 ** most expensive sub-scan. Add the smallest possible penalty 3386 ** (equivalent to multiplying the cost by 1.07) to ensure that 3387 ** this does not happen. Otherwise, for WHERE clauses such as the 3388 ** following where there is an index on "y": 3389 ** 3390 ** WHERE likelihood(x=?, 0.99) OR y=? 3391 ** 3392 ** the planner may elect to "OR" together a full-table scan and an 3393 ** index lookup. And other similarly odd results. */ 3394 pNew->rRun = sSum.a[i].rRun + 1; 3395 pNew->nOut = sSum.a[i].nOut; 3396 pNew->prereq = sSum.a[i].prereq; 3397 rc = whereLoopInsert(pBuilder, pNew); 3398 } 3399 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3400 } 3401 } 3402 return rc; 3403 } 3404 3405 /* 3406 ** Add all WhereLoop objects for all tables 3407 */ 3408 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3409 WhereInfo *pWInfo = pBuilder->pWInfo; 3410 Bitmask mPrereq = 0; 3411 Bitmask mPrior = 0; 3412 int iTab; 3413 SrcList *pTabList = pWInfo->pTabList; 3414 struct SrcList_item *pItem; 3415 struct SrcList_item *pEnd = &pTabList->a[pWInfo->nLevel]; 3416 sqlite3 *db = pWInfo->pParse->db; 3417 int rc = SQLITE_OK; 3418 WhereLoop *pNew; 3419 u8 priorJointype = 0; 3420 3421 /* Loop over the tables in the join, from left to right */ 3422 pNew = pBuilder->pNew; 3423 whereLoopInit(pNew); 3424 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3425 Bitmask mUnusable = 0; 3426 pNew->iTab = iTab; 3427 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3428 if( ((pItem->fg.jointype|priorJointype) & (JT_LEFT|JT_CROSS))!=0 ){ 3429 /* This condition is true when pItem is the FROM clause term on the 3430 ** right-hand-side of a LEFT or CROSS JOIN. */ 3431 mPrereq = mPrior; 3432 } 3433 priorJointype = pItem->fg.jointype; 3434 #ifndef SQLITE_OMIT_VIRTUALTABLE 3435 if( IsVirtual(pItem->pTab) ){ 3436 struct SrcList_item *p; 3437 for(p=&pItem[1]; p<pEnd; p++){ 3438 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3439 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3440 } 3441 } 3442 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 3443 }else 3444 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3445 { 3446 rc = whereLoopAddBtree(pBuilder, mPrereq); 3447 } 3448 if( rc==SQLITE_OK ){ 3449 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 3450 } 3451 mPrior |= pNew->maskSelf; 3452 if( rc || db->mallocFailed ) break; 3453 } 3454 3455 whereLoopClear(db, pNew); 3456 return rc; 3457 } 3458 3459 /* 3460 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 3461 ** parameters) to see if it outputs rows in the requested ORDER BY 3462 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3463 ** 3464 ** N>0: N terms of the ORDER BY clause are satisfied 3465 ** N==0: No terms of the ORDER BY clause are satisfied 3466 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3467 ** 3468 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3469 ** strict. With GROUP BY and DISTINCT the only requirement is that 3470 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3471 ** and DISTINCT do not require rows to appear in any particular order as long 3472 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3473 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3474 ** pOrderBy terms must be matched in strict left-to-right order. 3475 */ 3476 static i8 wherePathSatisfiesOrderBy( 3477 WhereInfo *pWInfo, /* The WHERE clause */ 3478 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3479 WherePath *pPath, /* The WherePath to check */ 3480 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 3481 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3482 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3483 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3484 ){ 3485 u8 revSet; /* True if rev is known */ 3486 u8 rev; /* Composite sort order */ 3487 u8 revIdx; /* Index sort order */ 3488 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 3489 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 3490 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 3491 u16 eqOpMask; /* Allowed equality operators */ 3492 u16 nKeyCol; /* Number of key columns in pIndex */ 3493 u16 nColumn; /* Total number of ordered columns in the index */ 3494 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 3495 int iLoop; /* Index of WhereLoop in pPath being processed */ 3496 int i, j; /* Loop counters */ 3497 int iCur; /* Cursor number for current WhereLoop */ 3498 int iColumn; /* A column number within table iCur */ 3499 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 3500 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3501 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 3502 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 3503 Index *pIndex; /* The index associated with pLoop */ 3504 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 3505 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 3506 Bitmask obDone; /* Mask of all ORDER BY terms */ 3507 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 3508 Bitmask ready; /* Mask of inner loops */ 3509 3510 /* 3511 ** We say the WhereLoop is "one-row" if it generates no more than one 3512 ** row of output. A WhereLoop is one-row if all of the following are true: 3513 ** (a) All index columns match with WHERE_COLUMN_EQ. 3514 ** (b) The index is unique 3515 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 3516 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 3517 ** 3518 ** We say the WhereLoop is "order-distinct" if the set of columns from 3519 ** that WhereLoop that are in the ORDER BY clause are different for every 3520 ** row of the WhereLoop. Every one-row WhereLoop is automatically 3521 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 3522 ** is not order-distinct. To be order-distinct is not quite the same as being 3523 ** UNIQUE since a UNIQUE column or index can have multiple rows that 3524 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 3525 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 3526 ** 3527 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 3528 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 3529 ** automatically order-distinct. 3530 */ 3531 3532 assert( pOrderBy!=0 ); 3533 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 3534 3535 nOrderBy = pOrderBy->nExpr; 3536 testcase( nOrderBy==BMS-1 ); 3537 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 3538 isOrderDistinct = 1; 3539 obDone = MASKBIT(nOrderBy)-1; 3540 orderDistinctMask = 0; 3541 ready = 0; 3542 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 3543 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) eqOpMask |= WO_IN; 3544 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 3545 if( iLoop>0 ) ready |= pLoop->maskSelf; 3546 if( iLoop<nLoop ){ 3547 pLoop = pPath->aLoop[iLoop]; 3548 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 3549 }else{ 3550 pLoop = pLast; 3551 } 3552 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 3553 if( pLoop->u.vtab.isOrdered ) obSat = obDone; 3554 break; 3555 }else{ 3556 pLoop->u.btree.nIdxCol = 0; 3557 } 3558 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 3559 3560 /* Mark off any ORDER BY term X that is a column in the table of 3561 ** the current loop for which there is term in the WHERE 3562 ** clause of the form X IS NULL or X=? that reference only outer 3563 ** loops. 3564 */ 3565 for(i=0; i<nOrderBy; i++){ 3566 if( MASKBIT(i) & obSat ) continue; 3567 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3568 if( pOBExpr->op!=TK_COLUMN ) continue; 3569 if( pOBExpr->iTable!=iCur ) continue; 3570 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 3571 ~ready, eqOpMask, 0); 3572 if( pTerm==0 ) continue; 3573 if( pTerm->eOperator==WO_IN ){ 3574 /* IN terms are only valid for sorting in the ORDER BY LIMIT 3575 ** optimization, and then only if they are actually used 3576 ** by the query plan */ 3577 assert( wctrlFlags & WHERE_ORDERBY_LIMIT ); 3578 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 3579 if( j>=pLoop->nLTerm ) continue; 3580 } 3581 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 3582 const char *z1, *z2; 3583 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3584 if( !pColl ) pColl = db->pDfltColl; 3585 z1 = pColl->zName; 3586 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pTerm->pExpr); 3587 if( !pColl ) pColl = db->pDfltColl; 3588 z2 = pColl->zName; 3589 if( sqlite3StrICmp(z1, z2)!=0 ) continue; 3590 testcase( pTerm->pExpr->op==TK_IS ); 3591 } 3592 obSat |= MASKBIT(i); 3593 } 3594 3595 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 3596 if( pLoop->wsFlags & WHERE_IPK ){ 3597 pIndex = 0; 3598 nKeyCol = 0; 3599 nColumn = 1; 3600 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 3601 return 0; 3602 }else{ 3603 nKeyCol = pIndex->nKeyCol; 3604 nColumn = pIndex->nColumn; 3605 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 3606 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 3607 || !HasRowid(pIndex->pTable)); 3608 isOrderDistinct = IsUniqueIndex(pIndex); 3609 } 3610 3611 /* Loop through all columns of the index and deal with the ones 3612 ** that are not constrained by == or IN. 3613 */ 3614 rev = revSet = 0; 3615 distinctColumns = 0; 3616 for(j=0; j<nColumn; j++){ 3617 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 3618 3619 assert( j>=pLoop->u.btree.nEq 3620 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 3621 ); 3622 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 3623 u16 eOp = pLoop->aLTerm[j]->eOperator; 3624 3625 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 3626 ** doing WHERE_ORDERBY_LIMIT processing). 3627 ** 3628 ** If the current term is a column of an ((?,?) IN (SELECT...)) 3629 ** expression for which the SELECT returns more than one column, 3630 ** check that it is the only column used by this loop. Otherwise, 3631 ** if it is one of two or more, none of the columns can be 3632 ** considered to match an ORDER BY term. */ 3633 if( (eOp & eqOpMask)!=0 ){ 3634 if( eOp & WO_ISNULL ){ 3635 testcase( isOrderDistinct ); 3636 isOrderDistinct = 0; 3637 } 3638 continue; 3639 }else if( ALWAYS(eOp & WO_IN) ){ 3640 /* ALWAYS() justification: eOp is an equality operator due to the 3641 ** j<pLoop->u.btree.nEq constraint above. Any equality other 3642 ** than WO_IN is captured by the previous "if". So this one 3643 ** always has to be WO_IN. */ 3644 Expr *pX = pLoop->aLTerm[j]->pExpr; 3645 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 3646 if( pLoop->aLTerm[i]->pExpr==pX ){ 3647 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 3648 bOnce = 0; 3649 break; 3650 } 3651 } 3652 } 3653 } 3654 3655 /* Get the column number in the table (iColumn) and sort order 3656 ** (revIdx) for the j-th column of the index. 3657 */ 3658 if( pIndex ){ 3659 iColumn = pIndex->aiColumn[j]; 3660 revIdx = pIndex->aSortOrder[j]; 3661 if( iColumn==pIndex->pTable->iPKey ) iColumn = -1; 3662 }else{ 3663 iColumn = XN_ROWID; 3664 revIdx = 0; 3665 } 3666 3667 /* An unconstrained column that might be NULL means that this 3668 ** WhereLoop is not well-ordered 3669 */ 3670 if( isOrderDistinct 3671 && iColumn>=0 3672 && j>=pLoop->u.btree.nEq 3673 && pIndex->pTable->aCol[iColumn].notNull==0 3674 ){ 3675 isOrderDistinct = 0; 3676 } 3677 3678 /* Find the ORDER BY term that corresponds to the j-th column 3679 ** of the index and mark that ORDER BY term off 3680 */ 3681 isMatch = 0; 3682 for(i=0; bOnce && i<nOrderBy; i++){ 3683 if( MASKBIT(i) & obSat ) continue; 3684 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3685 testcase( wctrlFlags & WHERE_GROUPBY ); 3686 testcase( wctrlFlags & WHERE_DISTINCTBY ); 3687 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 3688 if( iColumn>=(-1) ){ 3689 if( pOBExpr->op!=TK_COLUMN ) continue; 3690 if( pOBExpr->iTable!=iCur ) continue; 3691 if( pOBExpr->iColumn!=iColumn ) continue; 3692 }else{ 3693 if( sqlite3ExprCompare(0, 3694 pOBExpr,pIndex->aColExpr->a[j].pExpr,iCur) ){ 3695 continue; 3696 } 3697 } 3698 if( iColumn!=XN_ROWID ){ 3699 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3700 if( !pColl ) pColl = db->pDfltColl; 3701 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 3702 } 3703 pLoop->u.btree.nIdxCol = j+1; 3704 isMatch = 1; 3705 break; 3706 } 3707 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 3708 /* Make sure the sort order is compatible in an ORDER BY clause. 3709 ** Sort order is irrelevant for a GROUP BY clause. */ 3710 if( revSet ){ 3711 if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0; 3712 }else{ 3713 rev = revIdx ^ pOrderBy->a[i].sortOrder; 3714 if( rev ) *pRevMask |= MASKBIT(iLoop); 3715 revSet = 1; 3716 } 3717 } 3718 if( isMatch ){ 3719 if( iColumn==XN_ROWID ){ 3720 testcase( distinctColumns==0 ); 3721 distinctColumns = 1; 3722 } 3723 obSat |= MASKBIT(i); 3724 }else{ 3725 /* No match found */ 3726 if( j==0 || j<nKeyCol ){ 3727 testcase( isOrderDistinct!=0 ); 3728 isOrderDistinct = 0; 3729 } 3730 break; 3731 } 3732 } /* end Loop over all index columns */ 3733 if( distinctColumns ){ 3734 testcase( isOrderDistinct==0 ); 3735 isOrderDistinct = 1; 3736 } 3737 } /* end-if not one-row */ 3738 3739 /* Mark off any other ORDER BY terms that reference pLoop */ 3740 if( isOrderDistinct ){ 3741 orderDistinctMask |= pLoop->maskSelf; 3742 for(i=0; i<nOrderBy; i++){ 3743 Expr *p; 3744 Bitmask mTerm; 3745 if( MASKBIT(i) & obSat ) continue; 3746 p = pOrderBy->a[i].pExpr; 3747 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 3748 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 3749 if( (mTerm&~orderDistinctMask)==0 ){ 3750 obSat |= MASKBIT(i); 3751 } 3752 } 3753 } 3754 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 3755 if( obSat==obDone ) return (i8)nOrderBy; 3756 if( !isOrderDistinct ){ 3757 for(i=nOrderBy-1; i>0; i--){ 3758 Bitmask m = MASKBIT(i) - 1; 3759 if( (obSat&m)==m ) return i; 3760 } 3761 return 0; 3762 } 3763 return -1; 3764 } 3765 3766 3767 /* 3768 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 3769 ** the planner assumes that the specified pOrderBy list is actually a GROUP 3770 ** BY clause - and so any order that groups rows as required satisfies the 3771 ** request. 3772 ** 3773 ** Normally, in this case it is not possible for the caller to determine 3774 ** whether or not the rows are really being delivered in sorted order, or 3775 ** just in some other order that provides the required grouping. However, 3776 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 3777 ** this function may be called on the returned WhereInfo object. It returns 3778 ** true if the rows really will be sorted in the specified order, or false 3779 ** otherwise. 3780 ** 3781 ** For example, assuming: 3782 ** 3783 ** CREATE INDEX i1 ON t1(x, Y); 3784 ** 3785 ** then 3786 ** 3787 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 3788 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 3789 */ 3790 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 3791 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 3792 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 3793 return pWInfo->sorted; 3794 } 3795 3796 #ifdef WHERETRACE_ENABLED 3797 /* For debugging use only: */ 3798 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 3799 static char zName[65]; 3800 int i; 3801 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 3802 if( pLast ) zName[i++] = pLast->cId; 3803 zName[i] = 0; 3804 return zName; 3805 } 3806 #endif 3807 3808 /* 3809 ** Return the cost of sorting nRow rows, assuming that the keys have 3810 ** nOrderby columns and that the first nSorted columns are already in 3811 ** order. 3812 */ 3813 static LogEst whereSortingCost( 3814 WhereInfo *pWInfo, 3815 LogEst nRow, 3816 int nOrderBy, 3817 int nSorted 3818 ){ 3819 /* TUNING: Estimated cost of a full external sort, where N is 3820 ** the number of rows to sort is: 3821 ** 3822 ** cost = (3.0 * N * log(N)). 3823 ** 3824 ** Or, if the order-by clause has X terms but only the last Y 3825 ** terms are out of order, then block-sorting will reduce the 3826 ** sorting cost to: 3827 ** 3828 ** cost = (3.0 * N * log(N)) * (Y/X) 3829 ** 3830 ** The (Y/X) term is implemented using stack variable rScale 3831 ** below. */ 3832 LogEst rScale, rSortCost; 3833 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 3834 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 3835 rSortCost = nRow + rScale + 16; 3836 3837 /* Multiple by log(M) where M is the number of output rows. 3838 ** Use the LIMIT for M if it is smaller */ 3839 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 3840 nRow = pWInfo->iLimit; 3841 } 3842 rSortCost += estLog(nRow); 3843 return rSortCost; 3844 } 3845 3846 /* 3847 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 3848 ** attempts to find the lowest cost path that visits each WhereLoop 3849 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 3850 ** 3851 ** Assume that the total number of output rows that will need to be sorted 3852 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 3853 ** costs if nRowEst==0. 3854 ** 3855 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 3856 ** error occurs. 3857 */ 3858 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 3859 int mxChoice; /* Maximum number of simultaneous paths tracked */ 3860 int nLoop; /* Number of terms in the join */ 3861 Parse *pParse; /* Parsing context */ 3862 sqlite3 *db; /* The database connection */ 3863 int iLoop; /* Loop counter over the terms of the join */ 3864 int ii, jj; /* Loop counters */ 3865 int mxI = 0; /* Index of next entry to replace */ 3866 int nOrderBy; /* Number of ORDER BY clause terms */ 3867 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 3868 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 3869 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 3870 WherePath *aFrom; /* All nFrom paths at the previous level */ 3871 WherePath *aTo; /* The nTo best paths at the current level */ 3872 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 3873 WherePath *pTo; /* An element of aTo[] that we are working on */ 3874 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 3875 WhereLoop **pX; /* Used to divy up the pSpace memory */ 3876 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 3877 char *pSpace; /* Temporary memory used by this routine */ 3878 int nSpace; /* Bytes of space allocated at pSpace */ 3879 3880 pParse = pWInfo->pParse; 3881 db = pParse->db; 3882 nLoop = pWInfo->nLevel; 3883 /* TUNING: For simple queries, only the best path is tracked. 3884 ** For 2-way joins, the 5 best paths are followed. 3885 ** For joins of 3 or more tables, track the 10 best paths */ 3886 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 3887 assert( nLoop<=pWInfo->pTabList->nSrc ); 3888 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 3889 3890 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 3891 ** case the purpose of this call is to estimate the number of rows returned 3892 ** by the overall query. Once this estimate has been obtained, the caller 3893 ** will invoke this function a second time, passing the estimate as the 3894 ** nRowEst parameter. */ 3895 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 3896 nOrderBy = 0; 3897 }else{ 3898 nOrderBy = pWInfo->pOrderBy->nExpr; 3899 } 3900 3901 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 3902 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 3903 nSpace += sizeof(LogEst) * nOrderBy; 3904 pSpace = sqlite3DbMallocRawNN(db, nSpace); 3905 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 3906 aTo = (WherePath*)pSpace; 3907 aFrom = aTo+mxChoice; 3908 memset(aFrom, 0, sizeof(aFrom[0])); 3909 pX = (WhereLoop**)(aFrom+mxChoice); 3910 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 3911 pFrom->aLoop = pX; 3912 } 3913 if( nOrderBy ){ 3914 /* If there is an ORDER BY clause and it is not being ignored, set up 3915 ** space for the aSortCost[] array. Each element of the aSortCost array 3916 ** is either zero - meaning it has not yet been initialized - or the 3917 ** cost of sorting nRowEst rows of data where the first X terms of 3918 ** the ORDER BY clause are already in order, where X is the array 3919 ** index. */ 3920 aSortCost = (LogEst*)pX; 3921 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 3922 } 3923 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 3924 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 3925 3926 /* Seed the search with a single WherePath containing zero WhereLoops. 3927 ** 3928 ** TUNING: Do not let the number of iterations go above 28. If the cost 3929 ** of computing an automatic index is not paid back within the first 28 3930 ** rows, then do not use the automatic index. */ 3931 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 3932 nFrom = 1; 3933 assert( aFrom[0].isOrdered==0 ); 3934 if( nOrderBy ){ 3935 /* If nLoop is zero, then there are no FROM terms in the query. Since 3936 ** in this case the query may return a maximum of one row, the results 3937 ** are already in the requested order. Set isOrdered to nOrderBy to 3938 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 3939 ** -1, indicating that the result set may or may not be ordered, 3940 ** depending on the loops added to the current plan. */ 3941 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 3942 } 3943 3944 /* Compute successively longer WherePaths using the previous generation 3945 ** of WherePaths as the basis for the next. Keep track of the mxChoice 3946 ** best paths at each generation */ 3947 for(iLoop=0; iLoop<nLoop; iLoop++){ 3948 nTo = 0; 3949 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 3950 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 3951 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 3952 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 3953 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 3954 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 3955 Bitmask maskNew; /* Mask of src visited by (..) */ 3956 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 3957 3958 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 3959 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 3960 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<10 ){ 3961 /* Do not use an automatic index if the this loop is expected 3962 ** to run less than 2 times. */ 3963 assert( 10==sqlite3LogEst(2) ); 3964 continue; 3965 } 3966 /* At this point, pWLoop is a candidate to be the next loop. 3967 ** Compute its cost */ 3968 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 3969 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 3970 nOut = pFrom->nRow + pWLoop->nOut; 3971 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 3972 if( isOrdered<0 ){ 3973 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 3974 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 3975 iLoop, pWLoop, &revMask); 3976 }else{ 3977 revMask = pFrom->revLoop; 3978 } 3979 if( isOrdered>=0 && isOrdered<nOrderBy ){ 3980 if( aSortCost[isOrdered]==0 ){ 3981 aSortCost[isOrdered] = whereSortingCost( 3982 pWInfo, nRowEst, nOrderBy, isOrdered 3983 ); 3984 } 3985 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]); 3986 3987 WHERETRACE(0x002, 3988 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 3989 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 3990 rUnsorted, rCost)); 3991 }else{ 3992 rCost = rUnsorted; 3993 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 3994 } 3995 3996 /* Check to see if pWLoop should be added to the set of 3997 ** mxChoice best-so-far paths. 3998 ** 3999 ** First look for an existing path among best-so-far paths 4000 ** that covers the same set of loops and has the same isOrdered 4001 ** setting as the current path candidate. 4002 ** 4003 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 4004 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 4005 ** of legal values for isOrdered, -1..64. 4006 */ 4007 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4008 if( pTo->maskLoop==maskNew 4009 && ((pTo->isOrdered^isOrdered)&0x80)==0 4010 ){ 4011 testcase( jj==nTo-1 ); 4012 break; 4013 } 4014 } 4015 if( jj>=nTo ){ 4016 /* None of the existing best-so-far paths match the candidate. */ 4017 if( nTo>=mxChoice 4018 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4019 ){ 4020 /* The current candidate is no better than any of the mxChoice 4021 ** paths currently in the best-so-far buffer. So discard 4022 ** this candidate as not viable. */ 4023 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4024 if( sqlite3WhereTrace&0x4 ){ 4025 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4026 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4027 isOrdered>=0 ? isOrdered+'0' : '?'); 4028 } 4029 #endif 4030 continue; 4031 } 4032 /* If we reach this points it means that the new candidate path 4033 ** needs to be added to the set of best-so-far paths. */ 4034 if( nTo<mxChoice ){ 4035 /* Increase the size of the aTo set by one */ 4036 jj = nTo++; 4037 }else{ 4038 /* New path replaces the prior worst to keep count below mxChoice */ 4039 jj = mxI; 4040 } 4041 pTo = &aTo[jj]; 4042 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4043 if( sqlite3WhereTrace&0x4 ){ 4044 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4045 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4046 isOrdered>=0 ? isOrdered+'0' : '?'); 4047 } 4048 #endif 4049 }else{ 4050 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4051 ** same set of loops and has the same isOrdered setting as the 4052 ** candidate path. Check to see if the candidate should replace 4053 ** pTo or if the candidate should be skipped. 4054 ** 4055 ** The conditional is an expanded vector comparison equivalent to: 4056 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4057 */ 4058 if( pTo->rCost<rCost 4059 || (pTo->rCost==rCost 4060 && (pTo->nRow<nOut 4061 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4062 ) 4063 ) 4064 ){ 4065 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4066 if( sqlite3WhereTrace&0x4 ){ 4067 sqlite3DebugPrintf( 4068 "Skip %s cost=%-3d,%3d,%3d order=%c", 4069 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4070 isOrdered>=0 ? isOrdered+'0' : '?'); 4071 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4072 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4073 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4074 } 4075 #endif 4076 /* Discard the candidate path from further consideration */ 4077 testcase( pTo->rCost==rCost ); 4078 continue; 4079 } 4080 testcase( pTo->rCost==rCost+1 ); 4081 /* Control reaches here if the candidate path is better than the 4082 ** pTo path. Replace pTo with the candidate. */ 4083 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4084 if( sqlite3WhereTrace&0x4 ){ 4085 sqlite3DebugPrintf( 4086 "Update %s cost=%-3d,%3d,%3d order=%c", 4087 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4088 isOrdered>=0 ? isOrdered+'0' : '?'); 4089 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4090 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4091 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4092 } 4093 #endif 4094 } 4095 /* pWLoop is a winner. Add it to the set of best so far */ 4096 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4097 pTo->revLoop = revMask; 4098 pTo->nRow = nOut; 4099 pTo->rCost = rCost; 4100 pTo->rUnsorted = rUnsorted; 4101 pTo->isOrdered = isOrdered; 4102 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4103 pTo->aLoop[iLoop] = pWLoop; 4104 if( nTo>=mxChoice ){ 4105 mxI = 0; 4106 mxCost = aTo[0].rCost; 4107 mxUnsorted = aTo[0].nRow; 4108 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4109 if( pTo->rCost>mxCost 4110 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4111 ){ 4112 mxCost = pTo->rCost; 4113 mxUnsorted = pTo->rUnsorted; 4114 mxI = jj; 4115 } 4116 } 4117 } 4118 } 4119 } 4120 4121 #ifdef WHERETRACE_ENABLED /* >=2 */ 4122 if( sqlite3WhereTrace & 0x02 ){ 4123 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4124 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4125 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4126 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4127 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4128 if( pTo->isOrdered>0 ){ 4129 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4130 }else{ 4131 sqlite3DebugPrintf("\n"); 4132 } 4133 } 4134 } 4135 #endif 4136 4137 /* Swap the roles of aFrom and aTo for the next generation */ 4138 pFrom = aTo; 4139 aTo = aFrom; 4140 aFrom = pFrom; 4141 nFrom = nTo; 4142 } 4143 4144 if( nFrom==0 ){ 4145 sqlite3ErrorMsg(pParse, "no query solution"); 4146 sqlite3DbFreeNN(db, pSpace); 4147 return SQLITE_ERROR; 4148 } 4149 4150 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4151 pFrom = aFrom; 4152 for(ii=1; ii<nFrom; ii++){ 4153 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4154 } 4155 assert( pWInfo->nLevel==nLoop ); 4156 /* Load the lowest cost path into pWInfo */ 4157 for(iLoop=0; iLoop<nLoop; iLoop++){ 4158 WhereLevel *pLevel = pWInfo->a + iLoop; 4159 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4160 pLevel->iFrom = pWLoop->iTab; 4161 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4162 } 4163 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4164 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4165 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4166 && nRowEst 4167 ){ 4168 Bitmask notUsed; 4169 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4170 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4171 if( rc==pWInfo->pResultSet->nExpr ){ 4172 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4173 } 4174 } 4175 if( pWInfo->pOrderBy ){ 4176 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4177 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4178 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4179 } 4180 }else{ 4181 pWInfo->nOBSat = pFrom->isOrdered; 4182 pWInfo->revMask = pFrom->revLoop; 4183 if( pWInfo->nOBSat<=0 ){ 4184 pWInfo->nOBSat = 0; 4185 if( nLoop>0 ){ 4186 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4187 if( (wsFlags & WHERE_ONEROW)==0 4188 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4189 ){ 4190 Bitmask m = 0; 4191 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4192 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4193 testcase( wsFlags & WHERE_IPK ); 4194 testcase( wsFlags & WHERE_COLUMN_IN ); 4195 if( rc==pWInfo->pOrderBy->nExpr ){ 4196 pWInfo->bOrderedInnerLoop = 1; 4197 pWInfo->revMask = m; 4198 } 4199 } 4200 } 4201 } 4202 } 4203 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4204 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4205 ){ 4206 Bitmask revMask = 0; 4207 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4208 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4209 ); 4210 assert( pWInfo->sorted==0 ); 4211 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4212 pWInfo->sorted = 1; 4213 pWInfo->revMask = revMask; 4214 } 4215 } 4216 } 4217 4218 4219 pWInfo->nRowOut = pFrom->nRow; 4220 4221 /* Free temporary memory and return success */ 4222 sqlite3DbFreeNN(db, pSpace); 4223 return SQLITE_OK; 4224 } 4225 4226 /* 4227 ** Most queries use only a single table (they are not joins) and have 4228 ** simple == constraints against indexed fields. This routine attempts 4229 ** to plan those simple cases using much less ceremony than the 4230 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4231 ** times for the common case. 4232 ** 4233 ** Return non-zero on success, if this query can be handled by this 4234 ** no-frills query planner. Return zero if this query needs the 4235 ** general-purpose query planner. 4236 */ 4237 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4238 WhereInfo *pWInfo; 4239 struct SrcList_item *pItem; 4240 WhereClause *pWC; 4241 WhereTerm *pTerm; 4242 WhereLoop *pLoop; 4243 int iCur; 4244 int j; 4245 Table *pTab; 4246 Index *pIdx; 4247 4248 pWInfo = pBuilder->pWInfo; 4249 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4250 assert( pWInfo->pTabList->nSrc>=1 ); 4251 pItem = pWInfo->pTabList->a; 4252 pTab = pItem->pTab; 4253 if( IsVirtual(pTab) ) return 0; 4254 if( pItem->fg.isIndexedBy ) return 0; 4255 iCur = pItem->iCursor; 4256 pWC = &pWInfo->sWC; 4257 pLoop = pBuilder->pNew; 4258 pLoop->wsFlags = 0; 4259 pLoop->nSkip = 0; 4260 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0); 4261 if( pTerm ){ 4262 testcase( pTerm->eOperator & WO_IS ); 4263 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4264 pLoop->aLTerm[0] = pTerm; 4265 pLoop->nLTerm = 1; 4266 pLoop->u.btree.nEq = 1; 4267 /* TUNING: Cost of a rowid lookup is 10 */ 4268 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 4269 }else{ 4270 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 4271 int opMask; 4272 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 4273 if( !IsUniqueIndex(pIdx) 4274 || pIdx->pPartIdxWhere!=0 4275 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 4276 ) continue; 4277 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 4278 for(j=0; j<pIdx->nKeyCol; j++){ 4279 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx); 4280 if( pTerm==0 ) break; 4281 testcase( pTerm->eOperator & WO_IS ); 4282 pLoop->aLTerm[j] = pTerm; 4283 } 4284 if( j!=pIdx->nKeyCol ) continue; 4285 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 4286 if( pIdx->isCovering || (pItem->colUsed & ~columnsInIndex(pIdx))==0 ){ 4287 pLoop->wsFlags |= WHERE_IDX_ONLY; 4288 } 4289 pLoop->nLTerm = j; 4290 pLoop->u.btree.nEq = j; 4291 pLoop->u.btree.pIndex = pIdx; 4292 /* TUNING: Cost of a unique index lookup is 15 */ 4293 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 4294 break; 4295 } 4296 } 4297 if( pLoop->wsFlags ){ 4298 pLoop->nOut = (LogEst)1; 4299 pWInfo->a[0].pWLoop = pLoop; 4300 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 4301 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 4302 pWInfo->a[0].iTabCur = iCur; 4303 pWInfo->nRowOut = 1; 4304 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 4305 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 4306 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4307 } 4308 #ifdef SQLITE_DEBUG 4309 pLoop->cId = '0'; 4310 #endif 4311 return 1; 4312 } 4313 return 0; 4314 } 4315 4316 /* 4317 ** Helper function for exprIsDeterministic(). 4318 */ 4319 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 4320 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 4321 pWalker->eCode = 0; 4322 return WRC_Abort; 4323 } 4324 return WRC_Continue; 4325 } 4326 4327 /* 4328 ** Return true if the expression contains no non-deterministic SQL 4329 ** functions. Do not consider non-deterministic SQL functions that are 4330 ** part of sub-select statements. 4331 */ 4332 static int exprIsDeterministic(Expr *p){ 4333 Walker w; 4334 memset(&w, 0, sizeof(w)); 4335 w.eCode = 1; 4336 w.xExprCallback = exprNodeIsDeterministic; 4337 w.xSelectCallback = sqlite3SelectWalkFail; 4338 sqlite3WalkExpr(&w, p); 4339 return w.eCode; 4340 } 4341 4342 /* 4343 ** Generate the beginning of the loop used for WHERE clause processing. 4344 ** The return value is a pointer to an opaque structure that contains 4345 ** information needed to terminate the loop. Later, the calling routine 4346 ** should invoke sqlite3WhereEnd() with the return value of this function 4347 ** in order to complete the WHERE clause processing. 4348 ** 4349 ** If an error occurs, this routine returns NULL. 4350 ** 4351 ** The basic idea is to do a nested loop, one loop for each table in 4352 ** the FROM clause of a select. (INSERT and UPDATE statements are the 4353 ** same as a SELECT with only a single table in the FROM clause.) For 4354 ** example, if the SQL is this: 4355 ** 4356 ** SELECT * FROM t1, t2, t3 WHERE ...; 4357 ** 4358 ** Then the code generated is conceptually like the following: 4359 ** 4360 ** foreach row1 in t1 do \ Code generated 4361 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 4362 ** foreach row3 in t3 do / 4363 ** ... 4364 ** end \ Code generated 4365 ** end |-- by sqlite3WhereEnd() 4366 ** end / 4367 ** 4368 ** Note that the loops might not be nested in the order in which they 4369 ** appear in the FROM clause if a different order is better able to make 4370 ** use of indices. Note also that when the IN operator appears in 4371 ** the WHERE clause, it might result in additional nested loops for 4372 ** scanning through all values on the right-hand side of the IN. 4373 ** 4374 ** There are Btree cursors associated with each table. t1 uses cursor 4375 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 4376 ** And so forth. This routine generates code to open those VDBE cursors 4377 ** and sqlite3WhereEnd() generates the code to close them. 4378 ** 4379 ** The code that sqlite3WhereBegin() generates leaves the cursors named 4380 ** in pTabList pointing at their appropriate entries. The [...] code 4381 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 4382 ** data from the various tables of the loop. 4383 ** 4384 ** If the WHERE clause is empty, the foreach loops must each scan their 4385 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 4386 ** the tables have indices and there are terms in the WHERE clause that 4387 ** refer to those indices, a complete table scan can be avoided and the 4388 ** code will run much faster. Most of the work of this routine is checking 4389 ** to see if there are indices that can be used to speed up the loop. 4390 ** 4391 ** Terms of the WHERE clause are also used to limit which rows actually 4392 ** make it to the "..." in the middle of the loop. After each "foreach", 4393 ** terms of the WHERE clause that use only terms in that loop and outer 4394 ** loops are evaluated and if false a jump is made around all subsequent 4395 ** inner loops (or around the "..." if the test occurs within the inner- 4396 ** most loop) 4397 ** 4398 ** OUTER JOINS 4399 ** 4400 ** An outer join of tables t1 and t2 is conceptally coded as follows: 4401 ** 4402 ** foreach row1 in t1 do 4403 ** flag = 0 4404 ** foreach row2 in t2 do 4405 ** start: 4406 ** ... 4407 ** flag = 1 4408 ** end 4409 ** if flag==0 then 4410 ** move the row2 cursor to a null row 4411 ** goto start 4412 ** fi 4413 ** end 4414 ** 4415 ** ORDER BY CLAUSE PROCESSING 4416 ** 4417 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 4418 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 4419 ** if there is one. If there is no ORDER BY clause or if this routine 4420 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 4421 ** 4422 ** The iIdxCur parameter is the cursor number of an index. If 4423 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 4424 ** to use for OR clause processing. The WHERE clause should use this 4425 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 4426 ** the first cursor in an array of cursors for all indices. iIdxCur should 4427 ** be used to compute the appropriate cursor depending on which index is 4428 ** used. 4429 */ 4430 WhereInfo *sqlite3WhereBegin( 4431 Parse *pParse, /* The parser context */ 4432 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 4433 Expr *pWhere, /* The WHERE clause */ 4434 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 4435 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 4436 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 4437 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 4438 ** If WHERE_USE_LIMIT, then the limit amount */ 4439 ){ 4440 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4441 int nTabList; /* Number of elements in pTabList */ 4442 WhereInfo *pWInfo; /* Will become the return value of this function */ 4443 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4444 Bitmask notReady; /* Cursors that are not yet positioned */ 4445 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 4446 WhereMaskSet *pMaskSet; /* The expression mask set */ 4447 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 4448 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 4449 int ii; /* Loop counter */ 4450 sqlite3 *db; /* Database connection */ 4451 int rc; /* Return code */ 4452 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 4453 4454 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 4455 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 4456 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4457 )); 4458 4459 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 4460 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4461 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 4462 4463 /* Variable initialization */ 4464 db = pParse->db; 4465 memset(&sWLB, 0, sizeof(sWLB)); 4466 4467 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 4468 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 4469 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 4470 sWLB.pOrderBy = pOrderBy; 4471 4472 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 4473 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 4474 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 4475 wctrlFlags &= ~WHERE_WANT_DISTINCT; 4476 } 4477 4478 /* The number of tables in the FROM clause is limited by the number of 4479 ** bits in a Bitmask 4480 */ 4481 testcase( pTabList->nSrc==BMS ); 4482 if( pTabList->nSrc>BMS ){ 4483 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 4484 return 0; 4485 } 4486 4487 /* This function normally generates a nested loop for all tables in 4488 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 4489 ** only generate code for the first table in pTabList and assume that 4490 ** any cursors associated with subsequent tables are uninitialized. 4491 */ 4492 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 4493 4494 /* Allocate and initialize the WhereInfo structure that will become the 4495 ** return value. A single allocation is used to store the WhereInfo 4496 ** struct, the contents of WhereInfo.a[], the WhereClause structure 4497 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 4498 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 4499 ** some architectures. Hence the ROUND8() below. 4500 */ 4501 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 4502 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 4503 if( db->mallocFailed ){ 4504 sqlite3DbFree(db, pWInfo); 4505 pWInfo = 0; 4506 goto whereBeginError; 4507 } 4508 pWInfo->pParse = pParse; 4509 pWInfo->pTabList = pTabList; 4510 pWInfo->pOrderBy = pOrderBy; 4511 pWInfo->pWhere = pWhere; 4512 pWInfo->pResultSet = pResultSet; 4513 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 4514 pWInfo->nLevel = nTabList; 4515 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(v); 4516 pWInfo->wctrlFlags = wctrlFlags; 4517 pWInfo->iLimit = iAuxArg; 4518 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 4519 memset(&pWInfo->nOBSat, 0, 4520 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 4521 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 4522 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 4523 pMaskSet = &pWInfo->sMaskSet; 4524 sWLB.pWInfo = pWInfo; 4525 sWLB.pWC = &pWInfo->sWC; 4526 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 4527 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 4528 whereLoopInit(sWLB.pNew); 4529 #ifdef SQLITE_DEBUG 4530 sWLB.pNew->cId = '*'; 4531 #endif 4532 4533 /* Split the WHERE clause into separate subexpressions where each 4534 ** subexpression is separated by an AND operator. 4535 */ 4536 initMaskSet(pMaskSet); 4537 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 4538 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 4539 4540 /* Special case: No FROM clause 4541 */ 4542 if( nTabList==0 ){ 4543 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 4544 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4545 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4546 } 4547 }else{ 4548 /* Assign a bit from the bitmask to every term in the FROM clause. 4549 ** 4550 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 4551 ** 4552 ** The rule of the previous sentence ensures thta if X is the bitmask for 4553 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 4554 ** Knowing the bitmask for all tables to the left of a left join is 4555 ** important. Ticket #3015. 4556 ** 4557 ** Note that bitmasks are created for all pTabList->nSrc tables in 4558 ** pTabList, not just the first nTabList tables. nTabList is normally 4559 ** equal to pTabList->nSrc but might be shortened to 1 if the 4560 ** WHERE_OR_SUBCLAUSE flag is set. 4561 */ 4562 ii = 0; 4563 do{ 4564 createMask(pMaskSet, pTabList->a[ii].iCursor); 4565 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 4566 }while( (++ii)<pTabList->nSrc ); 4567 #ifdef SQLITE_DEBUG 4568 { 4569 Bitmask mx = 0; 4570 for(ii=0; ii<pTabList->nSrc; ii++){ 4571 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 4572 assert( m>=mx ); 4573 mx = m; 4574 } 4575 } 4576 #endif 4577 } 4578 4579 /* Analyze all of the subexpressions. */ 4580 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 4581 if( db->mallocFailed ) goto whereBeginError; 4582 4583 /* Special case: WHERE terms that do not refer to any tables in the join 4584 ** (constant expressions). Evaluate each such term, and jump over all the 4585 ** generated code if the result is not true. 4586 ** 4587 ** Do not do this if the expression contains non-deterministic functions 4588 ** that are not within a sub-select. This is not strictly required, but 4589 ** preserves SQLite's legacy behaviour in the following two cases: 4590 ** 4591 ** FROM ... WHERE random()>0; -- eval random() once per row 4592 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 4593 */ 4594 for(ii=0; ii<sWLB.pWC->nTerm; ii++){ 4595 WhereTerm *pT = &sWLB.pWC->a[ii]; 4596 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 4597 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 4598 pT->wtFlags |= TERM_CODED; 4599 } 4600 } 4601 4602 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4603 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 4604 /* The DISTINCT marking is pointless. Ignore it. */ 4605 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4606 }else if( pOrderBy==0 ){ 4607 /* Try to ORDER BY the result set to make distinct processing easier */ 4608 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 4609 pWInfo->pOrderBy = pResultSet; 4610 } 4611 } 4612 4613 /* Construct the WhereLoop objects */ 4614 #if defined(WHERETRACE_ENABLED) 4615 if( sqlite3WhereTrace & 0xffff ){ 4616 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 4617 if( wctrlFlags & WHERE_USE_LIMIT ){ 4618 sqlite3DebugPrintf(", limit: %d", iAuxArg); 4619 } 4620 sqlite3DebugPrintf(")\n"); 4621 } 4622 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 4623 sqlite3WhereClausePrint(sWLB.pWC); 4624 } 4625 #endif 4626 4627 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 4628 rc = whereLoopAddAll(&sWLB); 4629 if( rc ) goto whereBeginError; 4630 4631 #ifdef WHERETRACE_ENABLED 4632 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4633 WhereLoop *p; 4634 int i; 4635 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4636 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4637 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4638 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 4639 whereLoopPrint(p, sWLB.pWC); 4640 } 4641 } 4642 #endif 4643 4644 wherePathSolver(pWInfo, 0); 4645 if( db->mallocFailed ) goto whereBeginError; 4646 if( pWInfo->pOrderBy ){ 4647 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 4648 if( db->mallocFailed ) goto whereBeginError; 4649 } 4650 } 4651 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 4652 pWInfo->revMask = ALLBITS; 4653 } 4654 if( pParse->nErr || NEVER(db->mallocFailed) ){ 4655 goto whereBeginError; 4656 } 4657 #ifdef WHERETRACE_ENABLED 4658 if( sqlite3WhereTrace ){ 4659 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 4660 if( pWInfo->nOBSat>0 ){ 4661 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 4662 } 4663 switch( pWInfo->eDistinct ){ 4664 case WHERE_DISTINCT_UNIQUE: { 4665 sqlite3DebugPrintf(" DISTINCT=unique"); 4666 break; 4667 } 4668 case WHERE_DISTINCT_ORDERED: { 4669 sqlite3DebugPrintf(" DISTINCT=ordered"); 4670 break; 4671 } 4672 case WHERE_DISTINCT_UNORDERED: { 4673 sqlite3DebugPrintf(" DISTINCT=unordered"); 4674 break; 4675 } 4676 } 4677 sqlite3DebugPrintf("\n"); 4678 for(ii=0; ii<pWInfo->nLevel; ii++){ 4679 whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 4680 } 4681 } 4682 #endif 4683 /* Attempt to omit tables from the join that do not effect the result */ 4684 if( pWInfo->nLevel>=2 4685 && pResultSet!=0 4686 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 4687 ){ 4688 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet); 4689 if( sWLB.pOrderBy ){ 4690 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy); 4691 } 4692 while( pWInfo->nLevel>=2 ){ 4693 WhereTerm *pTerm, *pEnd; 4694 pLoop = pWInfo->a[pWInfo->nLevel-1].pWLoop; 4695 if( (pWInfo->pTabList->a[pLoop->iTab].fg.jointype & JT_LEFT)==0 ) break; 4696 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 4697 && (pLoop->wsFlags & WHERE_ONEROW)==0 4698 ){ 4699 break; 4700 } 4701 if( (tabUsed & pLoop->maskSelf)!=0 ) break; 4702 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; 4703 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4704 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 4705 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 4706 ){ 4707 break; 4708 } 4709 } 4710 if( pTerm<pEnd ) break; 4711 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 4712 pWInfo->nLevel--; 4713 nTabList--; 4714 } 4715 } 4716 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 4717 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 4718 4719 /* If the caller is an UPDATE or DELETE statement that is requesting 4720 ** to use a one-pass algorithm, determine if this is appropriate. 4721 */ 4722 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 4723 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 4724 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 4725 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 4726 if( bOnerow 4727 || ((wctrlFlags & WHERE_ONEPASS_MULTIROW)!=0 4728 && 0==(wsFlags & WHERE_VIRTUALTABLE)) 4729 ){ 4730 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 4731 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 4732 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 4733 bFordelete = OPFLAG_FORDELETE; 4734 } 4735 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 4736 } 4737 } 4738 } 4739 4740 /* Open all tables in the pTabList and any indices selected for 4741 ** searching those tables. 4742 */ 4743 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 4744 Table *pTab; /* Table to open */ 4745 int iDb; /* Index of database containing table/index */ 4746 struct SrcList_item *pTabItem; 4747 4748 pTabItem = &pTabList->a[pLevel->iFrom]; 4749 pTab = pTabItem->pTab; 4750 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 4751 pLoop = pLevel->pWLoop; 4752 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 4753 /* Do nothing */ 4754 }else 4755 #ifndef SQLITE_OMIT_VIRTUALTABLE 4756 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 4757 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 4758 int iCur = pTabItem->iCursor; 4759 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 4760 }else if( IsVirtual(pTab) ){ 4761 /* noop */ 4762 }else 4763 #endif 4764 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 4765 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 4766 int op = OP_OpenRead; 4767 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4768 op = OP_OpenWrite; 4769 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 4770 }; 4771 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 4772 assert( pTabItem->iCursor==pLevel->iTabCur ); 4773 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 4774 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 4775 if( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol<BMS && HasRowid(pTab) ){ 4776 Bitmask b = pTabItem->colUsed; 4777 int n = 0; 4778 for(; b; b=b>>1, n++){} 4779 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 4780 assert( n<=pTab->nCol ); 4781 } 4782 #ifdef SQLITE_ENABLE_CURSOR_HINTS 4783 if( pLoop->u.btree.pIndex!=0 ){ 4784 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 4785 }else 4786 #endif 4787 { 4788 sqlite3VdbeChangeP5(v, bFordelete); 4789 } 4790 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 4791 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 4792 (const u8*)&pTabItem->colUsed, P4_INT64); 4793 #endif 4794 }else{ 4795 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 4796 } 4797 if( pLoop->wsFlags & WHERE_INDEXED ){ 4798 Index *pIx = pLoop->u.btree.pIndex; 4799 int iIndexCur; 4800 int op = OP_OpenRead; 4801 /* iAuxArg is always set to a positive value if ONEPASS is possible */ 4802 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 4803 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 4804 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 4805 ){ 4806 /* This is one term of an OR-optimization using the PRIMARY KEY of a 4807 ** WITHOUT ROWID table. No need for a separate index */ 4808 iIndexCur = pLevel->iTabCur; 4809 op = 0; 4810 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4811 Index *pJ = pTabItem->pTab->pIndex; 4812 iIndexCur = iAuxArg; 4813 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 4814 while( ALWAYS(pJ) && pJ!=pIx ){ 4815 iIndexCur++; 4816 pJ = pJ->pNext; 4817 } 4818 op = OP_OpenWrite; 4819 pWInfo->aiCurOnePass[1] = iIndexCur; 4820 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 4821 iIndexCur = iAuxArg; 4822 op = OP_ReopenIdx; 4823 }else{ 4824 iIndexCur = pParse->nTab++; 4825 } 4826 pLevel->iIdxCur = iIndexCur; 4827 assert( pIx->pSchema==pTab->pSchema ); 4828 assert( iIndexCur>=0 ); 4829 if( op ){ 4830 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 4831 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 4832 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 4833 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 4834 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 4835 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 4836 ){ 4837 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); /* Hint to COMDB2 */ 4838 } 4839 VdbeComment((v, "%s", pIx->zName)); 4840 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 4841 { 4842 u64 colUsed = 0; 4843 int ii, jj; 4844 for(ii=0; ii<pIx->nColumn; ii++){ 4845 jj = pIx->aiColumn[ii]; 4846 if( jj<0 ) continue; 4847 if( jj>63 ) jj = 63; 4848 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 4849 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 4850 } 4851 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 4852 (u8*)&colUsed, P4_INT64); 4853 } 4854 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 4855 } 4856 } 4857 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 4858 } 4859 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 4860 if( db->mallocFailed ) goto whereBeginError; 4861 4862 /* Generate the code to do the search. Each iteration of the for 4863 ** loop below generates code for a single nested loop of the VM 4864 ** program. 4865 */ 4866 notReady = ~(Bitmask)0; 4867 for(ii=0; ii<nTabList; ii++){ 4868 int addrExplain; 4869 int wsFlags; 4870 pLevel = &pWInfo->a[ii]; 4871 wsFlags = pLevel->pWLoop->wsFlags; 4872 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 4873 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 4874 constructAutomaticIndex(pParse, &pWInfo->sWC, 4875 &pTabList->a[pLevel->iFrom], notReady, pLevel); 4876 if( db->mallocFailed ) goto whereBeginError; 4877 } 4878 #endif 4879 addrExplain = sqlite3WhereExplainOneScan( 4880 pParse, pTabList, pLevel, ii, pLevel->iFrom, wctrlFlags 4881 ); 4882 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 4883 notReady = sqlite3WhereCodeOneLoopStart(pWInfo, ii, notReady); 4884 pWInfo->iContinue = pLevel->addrCont; 4885 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 4886 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 4887 } 4888 } 4889 4890 /* Done. */ 4891 VdbeModuleComment((v, "Begin WHERE-core")); 4892 return pWInfo; 4893 4894 /* Jump here if malloc fails */ 4895 whereBeginError: 4896 if( pWInfo ){ 4897 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 4898 whereInfoFree(db, pWInfo); 4899 } 4900 return 0; 4901 } 4902 4903 /* 4904 ** Generate the end of the WHERE loop. See comments on 4905 ** sqlite3WhereBegin() for additional information. 4906 */ 4907 void sqlite3WhereEnd(WhereInfo *pWInfo){ 4908 Parse *pParse = pWInfo->pParse; 4909 Vdbe *v = pParse->pVdbe; 4910 int i; 4911 WhereLevel *pLevel; 4912 WhereLoop *pLoop; 4913 SrcList *pTabList = pWInfo->pTabList; 4914 sqlite3 *db = pParse->db; 4915 4916 /* Generate loop termination code. 4917 */ 4918 VdbeModuleComment((v, "End WHERE-core")); 4919 sqlite3ExprCacheClear(pParse); 4920 for(i=pWInfo->nLevel-1; i>=0; i--){ 4921 int addr; 4922 pLevel = &pWInfo->a[i]; 4923 pLoop = pLevel->pWLoop; 4924 if( pLevel->op!=OP_Noop ){ 4925 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 4926 int addrSeek = 0; 4927 Index *pIdx; 4928 int n; 4929 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 4930 && (pLoop->wsFlags & WHERE_INDEXED)!=0 4931 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 4932 && (n = pLoop->u.btree.nIdxCol)>0 4933 && pIdx->aiRowLogEst[n]>=36 4934 ){ 4935 int r1 = pParse->nMem+1; 4936 int j, op; 4937 for(j=0; j<n; j++){ 4938 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 4939 } 4940 pParse->nMem += n+1; 4941 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 4942 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 4943 VdbeCoverageIf(v, op==OP_SeekLT); 4944 VdbeCoverageIf(v, op==OP_SeekGT); 4945 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 4946 } 4947 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 4948 /* The common case: Advance to the next row */ 4949 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 4950 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 4951 sqlite3VdbeChangeP5(v, pLevel->p5); 4952 VdbeCoverage(v); 4953 VdbeCoverageIf(v, pLevel->op==OP_Next); 4954 VdbeCoverageIf(v, pLevel->op==OP_Prev); 4955 VdbeCoverageIf(v, pLevel->op==OP_VNext); 4956 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 4957 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 4958 #endif 4959 }else{ 4960 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 4961 } 4962 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 4963 struct InLoop *pIn; 4964 int j; 4965 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 4966 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 4967 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 4968 if( pIn->eEndLoopOp!=OP_Noop ){ 4969 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 4970 VdbeCoverage(v); 4971 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_PrevIfOpen); 4972 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_NextIfOpen); 4973 } 4974 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 4975 } 4976 } 4977 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 4978 if( pLevel->addrSkip ){ 4979 sqlite3VdbeGoto(v, pLevel->addrSkip); 4980 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 4981 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 4982 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 4983 } 4984 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 4985 if( pLevel->addrLikeRep ){ 4986 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 4987 pLevel->addrLikeRep); 4988 VdbeCoverage(v); 4989 } 4990 #endif 4991 if( pLevel->iLeftJoin ){ 4992 int ws = pLoop->wsFlags; 4993 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 4994 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 4995 if( (ws & WHERE_IDX_ONLY)==0 ){ 4996 sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor); 4997 } 4998 if( (ws & WHERE_INDEXED) 4999 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCovidx) 5000 ){ 5001 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 5002 } 5003 if( pLevel->op==OP_Return ){ 5004 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 5005 }else{ 5006 sqlite3VdbeGoto(v, pLevel->addrFirst); 5007 } 5008 sqlite3VdbeJumpHere(v, addr); 5009 } 5010 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 5011 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 5012 } 5013 5014 /* The "break" point is here, just past the end of the outer loop. 5015 ** Set it. 5016 */ 5017 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5018 5019 assert( pWInfo->nLevel<=pTabList->nSrc ); 5020 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5021 int k, last; 5022 VdbeOp *pOp; 5023 Index *pIdx = 0; 5024 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; 5025 Table *pTab = pTabItem->pTab; 5026 assert( pTab!=0 ); 5027 pLoop = pLevel->pWLoop; 5028 5029 /* For a co-routine, change all OP_Column references to the table of 5030 ** the co-routine into OP_Copy of result contained in a register. 5031 ** OP_Rowid becomes OP_Null. 5032 */ 5033 if( pTabItem->fg.viaCoroutine ){ 5034 testcase( pParse->db->mallocFailed ); 5035 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 5036 pTabItem->regResult, 0); 5037 continue; 5038 } 5039 5040 /* If this scan uses an index, make VDBE code substitutions to read data 5041 ** from the index instead of from the table where possible. In some cases 5042 ** this optimization prevents the table from ever being read, which can 5043 ** yield a significant performance boost. 5044 ** 5045 ** Calls to the code generator in between sqlite3WhereBegin and 5046 ** sqlite3WhereEnd will have created code that references the table 5047 ** directly. This loop scans all that code looking for opcodes 5048 ** that reference the table and converts them into opcodes that 5049 ** reference the index. 5050 */ 5051 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 5052 pIdx = pLoop->u.btree.pIndex; 5053 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 5054 pIdx = pLevel->u.pCovidx; 5055 } 5056 if( pIdx 5057 && (pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable)) 5058 && !db->mallocFailed 5059 ){ 5060 last = sqlite3VdbeCurrentAddr(v); 5061 k = pLevel->addrBody; 5062 pOp = sqlite3VdbeGetOp(v, k); 5063 for(; k<last; k++, pOp++){ 5064 if( pOp->p1!=pLevel->iTabCur ) continue; 5065 if( pOp->opcode==OP_Column ){ 5066 int x = pOp->p2; 5067 assert( pIdx->pTable==pTab ); 5068 if( !HasRowid(pTab) ){ 5069 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 5070 x = pPk->aiColumn[x]; 5071 assert( x>=0 ); 5072 } 5073 x = sqlite3ColumnOfIndex(pIdx, x); 5074 if( x>=0 ){ 5075 pOp->p2 = x; 5076 pOp->p1 = pLevel->iIdxCur; 5077 } 5078 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 5079 || pWInfo->eOnePass ); 5080 }else if( pOp->opcode==OP_Rowid ){ 5081 pOp->p1 = pLevel->iIdxCur; 5082 pOp->opcode = OP_IdxRowid; 5083 }else if( pOp->opcode==OP_IfNullRow ){ 5084 pOp->p1 = pLevel->iIdxCur; 5085 } 5086 } 5087 } 5088 } 5089 5090 /* Final cleanup 5091 */ 5092 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5093 whereInfoFree(db, pWInfo); 5094 return; 5095 } 5096