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 /* 23 ** Extra information appended to the end of sqlite3_index_info but not 24 ** visible to the xBestIndex function, at least not directly. The 25 ** sqlite3_vtab_collation() interface knows how to reach it, however. 26 ** 27 ** This object is not an API and can be changed from one release to the 28 ** next. As long as allocateIndexInfo() and sqlite3_vtab_collation() 29 ** agree on the structure, all will be well. 30 */ 31 typedef struct HiddenIndexInfo HiddenIndexInfo; 32 struct HiddenIndexInfo { 33 WhereClause *pWC; /* The Where clause being analyzed */ 34 Parse *pParse; /* The parsing context */ 35 }; 36 37 /* Forward declaration of methods */ 38 static int whereLoopResize(sqlite3*, WhereLoop*, int); 39 40 /* Test variable that can be set to enable WHERE tracing */ 41 #if defined(SQLITE_TEST) || defined(SQLITE_DEBUG) 42 /***/ int sqlite3WhereTrace = 0; 43 #endif 44 45 46 /* 47 ** Return the estimated number of output rows from a WHERE clause 48 */ 49 LogEst sqlite3WhereOutputRowCount(WhereInfo *pWInfo){ 50 return pWInfo->nRowOut; 51 } 52 53 /* 54 ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this 55 ** WHERE clause returns outputs for DISTINCT processing. 56 */ 57 int sqlite3WhereIsDistinct(WhereInfo *pWInfo){ 58 return pWInfo->eDistinct; 59 } 60 61 /* 62 ** Return TRUE if the WHERE clause returns rows in ORDER BY order. 63 ** Return FALSE if the output needs to be sorted. 64 */ 65 int sqlite3WhereIsOrdered(WhereInfo *pWInfo){ 66 return pWInfo->nOBSat; 67 } 68 69 /* 70 ** In the ORDER BY LIMIT optimization, if the inner-most loop is known 71 ** to emit rows in increasing order, and if the last row emitted by the 72 ** inner-most loop did not fit within the sorter, then we can skip all 73 ** subsequent rows for the current iteration of the inner loop (because they 74 ** will not fit in the sorter either) and continue with the second inner 75 ** loop - the loop immediately outside the inner-most. 76 ** 77 ** When a row does not fit in the sorter (because the sorter already 78 ** holds LIMIT+OFFSET rows that are smaller), then a jump is made to the 79 ** label returned by this function. 80 ** 81 ** If the ORDER BY LIMIT optimization applies, the jump destination should 82 ** be the continuation for the second-inner-most loop. If the ORDER BY 83 ** LIMIT optimization does not apply, then the jump destination should 84 ** be the continuation for the inner-most loop. 85 ** 86 ** It is always safe for this routine to return the continuation of the 87 ** inner-most loop, in the sense that a correct answer will result. 88 ** Returning the continuation the second inner loop is an optimization 89 ** that might make the code run a little faster, but should not change 90 ** the final answer. 91 */ 92 int sqlite3WhereOrderByLimitOptLabel(WhereInfo *pWInfo){ 93 WhereLevel *pInner; 94 if( !pWInfo->bOrderedInnerLoop ){ 95 /* The ORDER BY LIMIT optimization does not apply. Jump to the 96 ** continuation of the inner-most loop. */ 97 return pWInfo->iContinue; 98 } 99 pInner = &pWInfo->a[pWInfo->nLevel-1]; 100 assert( pInner->addrNxt!=0 ); 101 return pInner->addrNxt; 102 } 103 104 /* 105 ** Return the VDBE address or label to jump to in order to continue 106 ** immediately with the next row of a WHERE clause. 107 */ 108 int sqlite3WhereContinueLabel(WhereInfo *pWInfo){ 109 assert( pWInfo->iContinue!=0 ); 110 return pWInfo->iContinue; 111 } 112 113 /* 114 ** Return the VDBE address or label to jump to in order to break 115 ** out of a WHERE loop. 116 */ 117 int sqlite3WhereBreakLabel(WhereInfo *pWInfo){ 118 return pWInfo->iBreak; 119 } 120 121 /* 122 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to 123 ** operate directly on the rowis returned by a WHERE clause. Return 124 ** ONEPASS_SINGLE (1) if the statement can operation directly because only 125 ** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass 126 ** optimization can be used on multiple 127 ** 128 ** If the ONEPASS optimization is used (if this routine returns true) 129 ** then also write the indices of open cursors used by ONEPASS 130 ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data 131 ** table and iaCur[1] gets the cursor used by an auxiliary index. 132 ** Either value may be -1, indicating that cursor is not used. 133 ** Any cursors returned will have been opened for writing. 134 ** 135 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is 136 ** unable to use the ONEPASS optimization. 137 */ 138 int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){ 139 memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2); 140 #ifdef WHERETRACE_ENABLED 141 if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){ 142 sqlite3DebugPrintf("%s cursors: %d %d\n", 143 pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI", 144 aiCur[0], aiCur[1]); 145 } 146 #endif 147 return pWInfo->eOnePass; 148 } 149 150 /* 151 ** Move the content of pSrc into pDest 152 */ 153 static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){ 154 pDest->n = pSrc->n; 155 memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0])); 156 } 157 158 /* 159 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet. 160 ** 161 ** The new entry might overwrite an existing entry, or it might be 162 ** appended, or it might be discarded. Do whatever is the right thing 163 ** so that pSet keeps the N_OR_COST best entries seen so far. 164 */ 165 static int whereOrInsert( 166 WhereOrSet *pSet, /* The WhereOrSet to be updated */ 167 Bitmask prereq, /* Prerequisites of the new entry */ 168 LogEst rRun, /* Run-cost of the new entry */ 169 LogEst nOut /* Number of outputs for the new entry */ 170 ){ 171 u16 i; 172 WhereOrCost *p; 173 for(i=pSet->n, p=pSet->a; i>0; i--, p++){ 174 if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){ 175 goto whereOrInsert_done; 176 } 177 if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){ 178 return 0; 179 } 180 } 181 if( pSet->n<N_OR_COST ){ 182 p = &pSet->a[pSet->n++]; 183 p->nOut = nOut; 184 }else{ 185 p = pSet->a; 186 for(i=1; i<pSet->n; i++){ 187 if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i; 188 } 189 if( p->rRun<=rRun ) return 0; 190 } 191 whereOrInsert_done: 192 p->prereq = prereq; 193 p->rRun = rRun; 194 if( p->nOut>nOut ) p->nOut = nOut; 195 return 1; 196 } 197 198 /* 199 ** Return the bitmask for the given cursor number. Return 0 if 200 ** iCursor is not in the set. 201 */ 202 Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){ 203 int i; 204 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 ); 205 for(i=0; i<pMaskSet->n; i++){ 206 if( pMaskSet->ix[i]==iCursor ){ 207 return MASKBIT(i); 208 } 209 } 210 return 0; 211 } 212 213 /* 214 ** Create a new mask for cursor iCursor. 215 ** 216 ** There is one cursor per table in the FROM clause. The number of 217 ** tables in the FROM clause is limited by a test early in the 218 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] 219 ** array will never overflow. 220 */ 221 static void createMask(WhereMaskSet *pMaskSet, int iCursor){ 222 assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); 223 pMaskSet->ix[pMaskSet->n++] = iCursor; 224 } 225 226 /* 227 ** Advance to the next WhereTerm that matches according to the criteria 228 ** established when the pScan object was initialized by whereScanInit(). 229 ** Return NULL if there are no more matching WhereTerms. 230 */ 231 static WhereTerm *whereScanNext(WhereScan *pScan){ 232 int iCur; /* The cursor on the LHS of the term */ 233 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */ 234 Expr *pX; /* An expression being tested */ 235 WhereClause *pWC; /* Shorthand for pScan->pWC */ 236 WhereTerm *pTerm; /* The term being tested */ 237 int k = pScan->k; /* Where to start scanning */ 238 239 assert( pScan->iEquiv<=pScan->nEquiv ); 240 pWC = pScan->pWC; 241 while(1){ 242 iColumn = pScan->aiColumn[pScan->iEquiv-1]; 243 iCur = pScan->aiCur[pScan->iEquiv-1]; 244 assert( pWC!=0 ); 245 do{ 246 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){ 247 if( pTerm->leftCursor==iCur 248 && pTerm->u.leftColumn==iColumn 249 && (iColumn!=XN_EXPR 250 || sqlite3ExprCompareSkip(pTerm->pExpr->pLeft, 251 pScan->pIdxExpr,iCur)==0) 252 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin)) 253 ){ 254 if( (pTerm->eOperator & WO_EQUIV)!=0 255 && pScan->nEquiv<ArraySize(pScan->aiCur) 256 && (pX = sqlite3ExprSkipCollate(pTerm->pExpr->pRight))->op==TK_COLUMN 257 ){ 258 int j; 259 for(j=0; j<pScan->nEquiv; j++){ 260 if( pScan->aiCur[j]==pX->iTable 261 && pScan->aiColumn[j]==pX->iColumn ){ 262 break; 263 } 264 } 265 if( j==pScan->nEquiv ){ 266 pScan->aiCur[j] = pX->iTable; 267 pScan->aiColumn[j] = pX->iColumn; 268 pScan->nEquiv++; 269 } 270 } 271 if( (pTerm->eOperator & pScan->opMask)!=0 ){ 272 /* Verify the affinity and collating sequence match */ 273 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){ 274 CollSeq *pColl; 275 Parse *pParse = pWC->pWInfo->pParse; 276 pX = pTerm->pExpr; 277 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){ 278 continue; 279 } 280 assert(pX->pLeft); 281 pColl = sqlite3BinaryCompareCollSeq(pParse, 282 pX->pLeft, pX->pRight); 283 if( pColl==0 ) pColl = pParse->db->pDfltColl; 284 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){ 285 continue; 286 } 287 } 288 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0 289 && (pX = pTerm->pExpr->pRight)->op==TK_COLUMN 290 && pX->iTable==pScan->aiCur[0] 291 && pX->iColumn==pScan->aiColumn[0] 292 ){ 293 testcase( pTerm->eOperator & WO_IS ); 294 continue; 295 } 296 pScan->pWC = pWC; 297 pScan->k = k+1; 298 return pTerm; 299 } 300 } 301 } 302 pWC = pWC->pOuter; 303 k = 0; 304 }while( pWC!=0 ); 305 if( pScan->iEquiv>=pScan->nEquiv ) break; 306 pWC = pScan->pOrigWC; 307 k = 0; 308 pScan->iEquiv++; 309 } 310 return 0; 311 } 312 313 /* 314 ** This is whereScanInit() for the case of an index on an expression. 315 ** It is factored out into a separate tail-recursion subroutine so that 316 ** the normal whereScanInit() routine, which is a high-runner, does not 317 ** need to push registers onto the stack as part of its prologue. 318 */ 319 static SQLITE_NOINLINE WhereTerm *whereScanInitIndexExpr(WhereScan *pScan){ 320 pScan->idxaff = sqlite3ExprAffinity(pScan->pIdxExpr); 321 return whereScanNext(pScan); 322 } 323 324 /* 325 ** Initialize a WHERE clause scanner object. Return a pointer to the 326 ** first match. Return NULL if there are no matches. 327 ** 328 ** The scanner will be searching the WHERE clause pWC. It will look 329 ** for terms of the form "X <op> <expr>" where X is column iColumn of table 330 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx 331 ** must be one of the indexes of table iCur. 332 ** 333 ** The <op> must be one of the operators described by opMask. 334 ** 335 ** If the search is for X and the WHERE clause contains terms of the 336 ** form X=Y then this routine might also return terms of the form 337 ** "Y <op> <expr>". The number of levels of transitivity is limited, 338 ** but is enough to handle most commonly occurring SQL statements. 339 ** 340 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with 341 ** index pIdx. 342 */ 343 static WhereTerm *whereScanInit( 344 WhereScan *pScan, /* The WhereScan object being initialized */ 345 WhereClause *pWC, /* The WHERE clause to be scanned */ 346 int iCur, /* Cursor to scan for */ 347 int iColumn, /* Column to scan for */ 348 u32 opMask, /* Operator(s) to scan for */ 349 Index *pIdx /* Must be compatible with this index */ 350 ){ 351 pScan->pOrigWC = pWC; 352 pScan->pWC = pWC; 353 pScan->pIdxExpr = 0; 354 pScan->idxaff = 0; 355 pScan->zCollName = 0; 356 pScan->opMask = opMask; 357 pScan->k = 0; 358 pScan->aiCur[0] = iCur; 359 pScan->nEquiv = 1; 360 pScan->iEquiv = 1; 361 if( pIdx ){ 362 int j = iColumn; 363 iColumn = pIdx->aiColumn[j]; 364 if( iColumn==XN_EXPR ){ 365 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr; 366 pScan->zCollName = pIdx->azColl[j]; 367 pScan->aiColumn[0] = XN_EXPR; 368 return whereScanInitIndexExpr(pScan); 369 }else if( iColumn==pIdx->pTable->iPKey ){ 370 iColumn = XN_ROWID; 371 }else if( iColumn>=0 ){ 372 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; 373 pScan->zCollName = pIdx->azColl[j]; 374 } 375 }else if( iColumn==XN_EXPR ){ 376 return 0; 377 } 378 pScan->aiColumn[0] = iColumn; 379 return whereScanNext(pScan); 380 } 381 382 /* 383 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 384 ** where X is a reference to the iColumn of table iCur or of index pIdx 385 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by 386 ** the op parameter. Return a pointer to the term. Return 0 if not found. 387 ** 388 ** If pIdx!=0 then it must be one of the indexes of table iCur. 389 ** Search for terms matching the iColumn-th column of pIdx 390 ** rather than the iColumn-th column of table iCur. 391 ** 392 ** The term returned might by Y=<expr> if there is another constraint in 393 ** the WHERE clause that specifies that X=Y. Any such constraints will be 394 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 395 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11 396 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10 397 ** other equivalent values. Hence a search for X will return <expr> if X=A1 398 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>. 399 ** 400 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 401 ** then try for the one with no dependencies on <expr> - in other words where 402 ** <expr> is a constant expression of some kind. Only return entries of 403 ** the form "X <op> Y" where Y is a column in another table if no terms of 404 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 405 ** exist, try to return a term that does not use WO_EQUIV. 406 */ 407 WhereTerm *sqlite3WhereFindTerm( 408 WhereClause *pWC, /* The WHERE clause to be searched */ 409 int iCur, /* Cursor number of LHS */ 410 int iColumn, /* Column number of LHS */ 411 Bitmask notReady, /* RHS must not overlap with this mask */ 412 u32 op, /* Mask of WO_xx values describing operator */ 413 Index *pIdx /* Must be compatible with this index, if not NULL */ 414 ){ 415 WhereTerm *pResult = 0; 416 WhereTerm *p; 417 WhereScan scan; 418 419 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); 420 op &= WO_EQ|WO_IS; 421 while( p ){ 422 if( (p->prereqRight & notReady)==0 ){ 423 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){ 424 testcase( p->eOperator & WO_IS ); 425 return p; 426 } 427 if( pResult==0 ) pResult = p; 428 } 429 p = whereScanNext(&scan); 430 } 431 return pResult; 432 } 433 434 /* 435 ** This function searches pList for an entry that matches the iCol-th column 436 ** of index pIdx. 437 ** 438 ** If such an expression is found, its index in pList->a[] is returned. If 439 ** no expression is found, -1 is returned. 440 */ 441 static int findIndexCol( 442 Parse *pParse, /* Parse context */ 443 ExprList *pList, /* Expression list to search */ 444 int iBase, /* Cursor for table associated with pIdx */ 445 Index *pIdx, /* Index to match column of */ 446 int iCol /* Column of index to match */ 447 ){ 448 int i; 449 const char *zColl = pIdx->azColl[iCol]; 450 451 for(i=0; i<pList->nExpr; i++){ 452 Expr *p = sqlite3ExprSkipCollate(pList->a[i].pExpr); 453 if( p->op==TK_COLUMN 454 && p->iColumn==pIdx->aiColumn[iCol] 455 && p->iTable==iBase 456 ){ 457 CollSeq *pColl = sqlite3ExprNNCollSeq(pParse, pList->a[i].pExpr); 458 if( 0==sqlite3StrICmp(pColl->zName, zColl) ){ 459 return i; 460 } 461 } 462 } 463 464 return -1; 465 } 466 467 /* 468 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL 469 */ 470 static int indexColumnNotNull(Index *pIdx, int iCol){ 471 int j; 472 assert( pIdx!=0 ); 473 assert( iCol>=0 && iCol<pIdx->nColumn ); 474 j = pIdx->aiColumn[iCol]; 475 if( j>=0 ){ 476 return pIdx->pTable->aCol[j].notNull; 477 }else if( j==(-1) ){ 478 return 1; 479 }else{ 480 assert( j==(-2) ); 481 return 0; /* Assume an indexed expression can always yield a NULL */ 482 483 } 484 } 485 486 /* 487 ** Return true if the DISTINCT expression-list passed as the third argument 488 ** is redundant. 489 ** 490 ** A DISTINCT list is redundant if any subset of the columns in the 491 ** DISTINCT list are collectively unique and individually non-null. 492 */ 493 static int isDistinctRedundant( 494 Parse *pParse, /* Parsing context */ 495 SrcList *pTabList, /* The FROM clause */ 496 WhereClause *pWC, /* The WHERE clause */ 497 ExprList *pDistinct /* The result set that needs to be DISTINCT */ 498 ){ 499 Table *pTab; 500 Index *pIdx; 501 int i; 502 int iBase; 503 504 /* If there is more than one table or sub-select in the FROM clause of 505 ** this query, then it will not be possible to show that the DISTINCT 506 ** clause is redundant. */ 507 if( pTabList->nSrc!=1 ) return 0; 508 iBase = pTabList->a[0].iCursor; 509 pTab = pTabList->a[0].pTab; 510 511 /* If any of the expressions is an IPK column on table iBase, then return 512 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 513 ** current SELECT is a correlated sub-query. 514 */ 515 for(i=0; i<pDistinct->nExpr; i++){ 516 Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr); 517 if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1; 518 } 519 520 /* Loop through all indices on the table, checking each to see if it makes 521 ** the DISTINCT qualifier redundant. It does so if: 522 ** 523 ** 1. The index is itself UNIQUE, and 524 ** 525 ** 2. All of the columns in the index are either part of the pDistinct 526 ** list, or else the WHERE clause contains a term of the form "col=X", 527 ** where X is a constant value. The collation sequences of the 528 ** comparison and select-list expressions must match those of the index. 529 ** 530 ** 3. All of those index columns for which the WHERE clause does not 531 ** contain a "col=X" term are subject to a NOT NULL constraint. 532 */ 533 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 534 if( !IsUniqueIndex(pIdx) ) continue; 535 for(i=0; i<pIdx->nKeyCol; i++){ 536 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){ 537 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break; 538 if( indexColumnNotNull(pIdx, i)==0 ) break; 539 } 540 } 541 if( i==pIdx->nKeyCol ){ 542 /* This index implies that the DISTINCT qualifier is redundant. */ 543 return 1; 544 } 545 } 546 547 return 0; 548 } 549 550 551 /* 552 ** Estimate the logarithm of the input value to base 2. 553 */ 554 static LogEst estLog(LogEst N){ 555 return N<=10 ? 0 : sqlite3LogEst(N) - 33; 556 } 557 558 /* 559 ** Convert OP_Column opcodes to OP_Copy in previously generated code. 560 ** 561 ** This routine runs over generated VDBE code and translates OP_Column 562 ** opcodes into OP_Copy when the table is being accessed via co-routine 563 ** instead of via table lookup. 564 ** 565 ** If the iAutoidxCur is not zero, then any OP_Rowid instructions on 566 ** cursor iTabCur are transformed into OP_Sequence opcode for the 567 ** iAutoidxCur cursor, in order to generate unique rowids for the 568 ** automatic index being generated. 569 */ 570 static void translateColumnToCopy( 571 Parse *pParse, /* Parsing context */ 572 int iStart, /* Translate from this opcode to the end */ 573 int iTabCur, /* OP_Column/OP_Rowid references to this table */ 574 int iRegister, /* The first column is in this register */ 575 int iAutoidxCur /* If non-zero, cursor of autoindex being generated */ 576 ){ 577 Vdbe *v = pParse->pVdbe; 578 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart); 579 int iEnd = sqlite3VdbeCurrentAddr(v); 580 if( pParse->db->mallocFailed ) return; 581 for(; iStart<iEnd; iStart++, pOp++){ 582 if( pOp->p1!=iTabCur ) continue; 583 if( pOp->opcode==OP_Column ){ 584 pOp->opcode = OP_Copy; 585 pOp->p1 = pOp->p2 + iRegister; 586 pOp->p2 = pOp->p3; 587 pOp->p3 = 0; 588 }else if( pOp->opcode==OP_Rowid ){ 589 if( iAutoidxCur ){ 590 pOp->opcode = OP_Sequence; 591 pOp->p1 = iAutoidxCur; 592 }else{ 593 pOp->opcode = OP_Null; 594 pOp->p1 = 0; 595 pOp->p3 = 0; 596 } 597 } 598 } 599 } 600 601 /* 602 ** Two routines for printing the content of an sqlite3_index_info 603 ** structure. Used for testing and debugging only. If neither 604 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 605 ** are no-ops. 606 */ 607 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) 608 static void TRACE_IDX_INPUTS(sqlite3_index_info *p){ 609 int i; 610 if( !sqlite3WhereTrace ) return; 611 for(i=0; i<p->nConstraint; i++){ 612 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", 613 i, 614 p->aConstraint[i].iColumn, 615 p->aConstraint[i].iTermOffset, 616 p->aConstraint[i].op, 617 p->aConstraint[i].usable); 618 } 619 for(i=0; i<p->nOrderBy; i++){ 620 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 621 i, 622 p->aOrderBy[i].iColumn, 623 p->aOrderBy[i].desc); 624 } 625 } 626 static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){ 627 int i; 628 if( !sqlite3WhereTrace ) return; 629 for(i=0; i<p->nConstraint; i++){ 630 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 631 i, 632 p->aConstraintUsage[i].argvIndex, 633 p->aConstraintUsage[i].omit); 634 } 635 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 636 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 637 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 638 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 639 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); 640 } 641 #else 642 #define TRACE_IDX_INPUTS(A) 643 #define TRACE_IDX_OUTPUTS(A) 644 #endif 645 646 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 647 /* 648 ** Return TRUE if the WHERE clause term pTerm is of a form where it 649 ** could be used with an index to access pSrc, assuming an appropriate 650 ** index existed. 651 */ 652 static int termCanDriveIndex( 653 WhereTerm *pTerm, /* WHERE clause term to check */ 654 struct SrcList_item *pSrc, /* Table we are trying to access */ 655 Bitmask notReady /* Tables in outer loops of the join */ 656 ){ 657 char aff; 658 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 659 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0; 660 if( (pSrc->fg.jointype & JT_LEFT) 661 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 662 && (pTerm->eOperator & WO_IS) 663 ){ 664 /* Cannot use an IS term from the WHERE clause as an index driver for 665 ** the RHS of a LEFT JOIN. Such a term can only be used if it is from 666 ** the ON clause. */ 667 return 0; 668 } 669 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 670 if( pTerm->u.leftColumn<0 ) return 0; 671 aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity; 672 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 673 testcase( pTerm->pExpr->op==TK_IS ); 674 return 1; 675 } 676 #endif 677 678 679 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 680 /* 681 ** Generate code to construct the Index object for an automatic index 682 ** and to set up the WhereLevel object pLevel so that the code generator 683 ** makes use of the automatic index. 684 */ 685 static void constructAutomaticIndex( 686 Parse *pParse, /* The parsing context */ 687 WhereClause *pWC, /* The WHERE clause */ 688 struct SrcList_item *pSrc, /* The FROM clause term to get the next index */ 689 Bitmask notReady, /* Mask of cursors that are not available */ 690 WhereLevel *pLevel /* Write new index here */ 691 ){ 692 int nKeyCol; /* Number of columns in the constructed index */ 693 WhereTerm *pTerm; /* A single term of the WHERE clause */ 694 WhereTerm *pWCEnd; /* End of pWC->a[] */ 695 Index *pIdx; /* Object describing the transient index */ 696 Vdbe *v; /* Prepared statement under construction */ 697 int addrInit; /* Address of the initialization bypass jump */ 698 Table *pTable; /* The table being indexed */ 699 int addrTop; /* Top of the index fill loop */ 700 int regRecord; /* Register holding an index record */ 701 int n; /* Column counter */ 702 int i; /* Loop counter */ 703 int mxBitCol; /* Maximum column in pSrc->colUsed */ 704 CollSeq *pColl; /* Collating sequence to on a column */ 705 WhereLoop *pLoop; /* The Loop object */ 706 char *zNotUsed; /* Extra space on the end of pIdx */ 707 Bitmask idxCols; /* Bitmap of columns used for indexing */ 708 Bitmask extraCols; /* Bitmap of additional columns */ 709 u8 sentWarning = 0; /* True if a warnning has been issued */ 710 Expr *pPartial = 0; /* Partial Index Expression */ 711 int iContinue = 0; /* Jump here to skip excluded rows */ 712 struct SrcList_item *pTabItem; /* FROM clause term being indexed */ 713 int addrCounter = 0; /* Address where integer counter is initialized */ 714 int regBase; /* Array of registers where record is assembled */ 715 716 /* Generate code to skip over the creation and initialization of the 717 ** transient index on 2nd and subsequent iterations of the loop. */ 718 v = pParse->pVdbe; 719 assert( v!=0 ); 720 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 721 722 /* Count the number of columns that will be added to the index 723 ** and used to match WHERE clause constraints */ 724 nKeyCol = 0; 725 pTable = pSrc->pTab; 726 pWCEnd = &pWC->a[pWC->nTerm]; 727 pLoop = pLevel->pWLoop; 728 idxCols = 0; 729 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 730 Expr *pExpr = pTerm->pExpr; 731 assert( !ExprHasProperty(pExpr, EP_FromJoin) /* prereq always non-zero */ 732 || pExpr->iRightJoinTable!=pSrc->iCursor /* for the right-hand */ 733 || pLoop->prereq!=0 ); /* table of a LEFT JOIN */ 734 if( pLoop->prereq==0 735 && (pTerm->wtFlags & TERM_VIRTUAL)==0 736 && !ExprHasProperty(pExpr, EP_FromJoin) 737 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) ){ 738 pPartial = sqlite3ExprAnd(pParse, pPartial, 739 sqlite3ExprDup(pParse->db, pExpr, 0)); 740 } 741 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 742 int iCol = pTerm->u.leftColumn; 743 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 744 testcase( iCol==BMS ); 745 testcase( iCol==BMS-1 ); 746 if( !sentWarning ){ 747 sqlite3_log(SQLITE_WARNING_AUTOINDEX, 748 "automatic index on %s(%s)", pTable->zName, 749 pTable->aCol[iCol].zName); 750 sentWarning = 1; 751 } 752 if( (idxCols & cMask)==0 ){ 753 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ 754 goto end_auto_index_create; 755 } 756 pLoop->aLTerm[nKeyCol++] = pTerm; 757 idxCols |= cMask; 758 } 759 } 760 } 761 assert( nKeyCol>0 ); 762 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; 763 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED 764 | WHERE_AUTO_INDEX; 765 766 /* Count the number of additional columns needed to create a 767 ** covering index. A "covering index" is an index that contains all 768 ** columns that are needed by the query. With a covering index, the 769 ** original table never needs to be accessed. Automatic indices must 770 ** be a covering index because the index will not be updated if the 771 ** original table changes and the index and table cannot both be used 772 ** if they go out of sync. 773 */ 774 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); 775 mxBitCol = MIN(BMS-1,pTable->nCol); 776 testcase( pTable->nCol==BMS-1 ); 777 testcase( pTable->nCol==BMS-2 ); 778 for(i=0; i<mxBitCol; i++){ 779 if( extraCols & MASKBIT(i) ) nKeyCol++; 780 } 781 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 782 nKeyCol += pTable->nCol - BMS + 1; 783 } 784 785 /* Construct the Index object to describe this index */ 786 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); 787 if( pIdx==0 ) goto end_auto_index_create; 788 pLoop->u.btree.pIndex = pIdx; 789 pIdx->zName = "auto-index"; 790 pIdx->pTable = pTable; 791 n = 0; 792 idxCols = 0; 793 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 794 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 795 int iCol = pTerm->u.leftColumn; 796 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 797 testcase( iCol==BMS-1 ); 798 testcase( iCol==BMS ); 799 if( (idxCols & cMask)==0 ){ 800 Expr *pX = pTerm->pExpr; 801 idxCols |= cMask; 802 pIdx->aiColumn[n] = pTerm->u.leftColumn; 803 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); 804 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY; 805 n++; 806 } 807 } 808 } 809 assert( (u32)n==pLoop->u.btree.nEq ); 810 811 /* Add additional columns needed to make the automatic index into 812 ** a covering index */ 813 for(i=0; i<mxBitCol; i++){ 814 if( extraCols & MASKBIT(i) ){ 815 pIdx->aiColumn[n] = i; 816 pIdx->azColl[n] = sqlite3StrBINARY; 817 n++; 818 } 819 } 820 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 821 for(i=BMS-1; i<pTable->nCol; i++){ 822 pIdx->aiColumn[n] = i; 823 pIdx->azColl[n] = sqlite3StrBINARY; 824 n++; 825 } 826 } 827 assert( n==nKeyCol ); 828 pIdx->aiColumn[n] = XN_ROWID; 829 pIdx->azColl[n] = sqlite3StrBINARY; 830 831 /* Create the automatic index */ 832 assert( pLevel->iIdxCur>=0 ); 833 pLevel->iIdxCur = pParse->nTab++; 834 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); 835 sqlite3VdbeSetP4KeyInfo(pParse, pIdx); 836 VdbeComment((v, "for %s", pTable->zName)); 837 838 /* Fill the automatic index with content */ 839 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom]; 840 if( pTabItem->fg.viaCoroutine ){ 841 int regYield = pTabItem->regReturn; 842 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0); 843 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); 844 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 845 VdbeCoverage(v); 846 VdbeComment((v, "next row of %s", pTabItem->pTab->zName)); 847 }else{ 848 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); 849 } 850 if( pPartial ){ 851 iContinue = sqlite3VdbeMakeLabel(pParse); 852 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); 853 pLoop->wsFlags |= WHERE_PARTIALIDX; 854 } 855 regRecord = sqlite3GetTempReg(pParse); 856 regBase = sqlite3GenerateIndexKey( 857 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0 858 ); 859 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 860 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 861 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); 862 if( pTabItem->fg.viaCoroutine ){ 863 sqlite3VdbeChangeP2(v, addrCounter, regBase+n); 864 testcase( pParse->db->mallocFailed ); 865 assert( pLevel->iIdxCur>0 ); 866 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur, 867 pTabItem->regResult, pLevel->iIdxCur); 868 sqlite3VdbeGoto(v, addrTop); 869 pTabItem->fg.viaCoroutine = 0; 870 }else{ 871 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); 872 } 873 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 874 sqlite3VdbeJumpHere(v, addrTop); 875 sqlite3ReleaseTempReg(pParse, regRecord); 876 877 /* Jump here when skipping the initialization */ 878 sqlite3VdbeJumpHere(v, addrInit); 879 880 end_auto_index_create: 881 sqlite3ExprDelete(pParse->db, pPartial); 882 } 883 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 884 885 #ifndef SQLITE_OMIT_VIRTUALTABLE 886 /* 887 ** Allocate and populate an sqlite3_index_info structure. It is the 888 ** responsibility of the caller to eventually release the structure 889 ** by passing the pointer returned by this function to sqlite3_free(). 890 */ 891 static sqlite3_index_info *allocateIndexInfo( 892 Parse *pParse, /* The parsing context */ 893 WhereClause *pWC, /* The WHERE clause being analyzed */ 894 Bitmask mUnusable, /* Ignore terms with these prereqs */ 895 struct SrcList_item *pSrc, /* The FROM clause term that is the vtab */ 896 ExprList *pOrderBy, /* The ORDER BY clause */ 897 u16 *pmNoOmit /* Mask of terms not to omit */ 898 ){ 899 int i, j; 900 int nTerm; 901 struct sqlite3_index_constraint *pIdxCons; 902 struct sqlite3_index_orderby *pIdxOrderBy; 903 struct sqlite3_index_constraint_usage *pUsage; 904 struct HiddenIndexInfo *pHidden; 905 WhereTerm *pTerm; 906 int nOrderBy; 907 sqlite3_index_info *pIdxInfo; 908 u16 mNoOmit = 0; 909 910 /* Count the number of possible WHERE clause constraints referring 911 ** to this virtual table */ 912 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 913 if( pTerm->leftCursor != pSrc->iCursor ) continue; 914 if( pTerm->prereqRight & mUnusable ) continue; 915 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 916 testcase( pTerm->eOperator & WO_IN ); 917 testcase( pTerm->eOperator & WO_ISNULL ); 918 testcase( pTerm->eOperator & WO_IS ); 919 testcase( pTerm->eOperator & WO_ALL ); 920 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 921 if( pTerm->wtFlags & TERM_VNULL ) continue; 922 assert( pTerm->u.leftColumn>=(-1) ); 923 nTerm++; 924 } 925 926 /* If the ORDER BY clause contains only columns in the current 927 ** virtual table then allocate space for the aOrderBy part of 928 ** the sqlite3_index_info structure. 929 */ 930 nOrderBy = 0; 931 if( pOrderBy ){ 932 int n = pOrderBy->nExpr; 933 for(i=0; i<n; i++){ 934 Expr *pExpr = pOrderBy->a[i].pExpr; 935 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; 936 } 937 if( i==n){ 938 nOrderBy = n; 939 } 940 } 941 942 /* Allocate the sqlite3_index_info structure 943 */ 944 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 945 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 946 + sizeof(*pIdxOrderBy)*nOrderBy + sizeof(*pHidden) ); 947 if( pIdxInfo==0 ){ 948 sqlite3ErrorMsg(pParse, "out of memory"); 949 return 0; 950 } 951 952 /* Initialize the structure. The sqlite3_index_info structure contains 953 ** many fields that are declared "const" to prevent xBestIndex from 954 ** changing them. We have to do some funky casting in order to 955 ** initialize those fields. 956 */ 957 pHidden = (struct HiddenIndexInfo*)&pIdxInfo[1]; 958 pIdxCons = (struct sqlite3_index_constraint*)&pHidden[1]; 959 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 960 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 961 *(int*)&pIdxInfo->nConstraint = nTerm; 962 *(int*)&pIdxInfo->nOrderBy = nOrderBy; 963 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons; 964 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy; 965 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage = 966 pUsage; 967 968 pHidden->pWC = pWC; 969 pHidden->pParse = pParse; 970 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 971 u16 op; 972 if( pTerm->leftCursor != pSrc->iCursor ) continue; 973 if( pTerm->prereqRight & mUnusable ) continue; 974 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 975 testcase( pTerm->eOperator & WO_IN ); 976 testcase( pTerm->eOperator & WO_IS ); 977 testcase( pTerm->eOperator & WO_ISNULL ); 978 testcase( pTerm->eOperator & WO_ALL ); 979 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 980 if( pTerm->wtFlags & TERM_VNULL ) continue; 981 if( (pSrc->fg.jointype & JT_LEFT)!=0 982 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 983 && (pTerm->eOperator & (WO_IS|WO_ISNULL)) 984 ){ 985 /* An "IS" term in the WHERE clause where the virtual table is the rhs 986 ** of a LEFT JOIN. Do not pass this term to the virtual table 987 ** implementation, as this can lead to incorrect results from SQL such 988 ** as: 989 ** 990 ** "LEFT JOIN vtab WHERE vtab.col IS NULL" */ 991 testcase( pTerm->eOperator & WO_ISNULL ); 992 testcase( pTerm->eOperator & WO_IS ); 993 continue; 994 } 995 assert( pTerm->u.leftColumn>=(-1) ); 996 pIdxCons[j].iColumn = pTerm->u.leftColumn; 997 pIdxCons[j].iTermOffset = i; 998 op = pTerm->eOperator & WO_ALL; 999 if( op==WO_IN ) op = WO_EQ; 1000 if( op==WO_AUX ){ 1001 pIdxCons[j].op = pTerm->eMatchOp; 1002 }else if( op & (WO_ISNULL|WO_IS) ){ 1003 if( op==WO_ISNULL ){ 1004 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_ISNULL; 1005 }else{ 1006 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_IS; 1007 } 1008 }else{ 1009 pIdxCons[j].op = (u8)op; 1010 /* The direct assignment in the previous line is possible only because 1011 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 1012 ** following asserts verify this fact. */ 1013 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 1014 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 1015 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 1016 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 1017 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 1018 assert( pTerm->eOperator&(WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_AUX) ); 1019 1020 if( op & (WO_LT|WO_LE|WO_GT|WO_GE) 1021 && sqlite3ExprIsVector(pTerm->pExpr->pRight) 1022 ){ 1023 if( i<16 ) mNoOmit |= (1 << i); 1024 if( op==WO_LT ) pIdxCons[j].op = WO_LE; 1025 if( op==WO_GT ) pIdxCons[j].op = WO_GE; 1026 } 1027 } 1028 1029 j++; 1030 } 1031 for(i=0; i<nOrderBy; i++){ 1032 Expr *pExpr = pOrderBy->a[i].pExpr; 1033 pIdxOrderBy[i].iColumn = pExpr->iColumn; 1034 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder; 1035 } 1036 1037 *pmNoOmit = mNoOmit; 1038 return pIdxInfo; 1039 } 1040 1041 /* 1042 ** The table object reference passed as the second argument to this function 1043 ** must represent a virtual table. This function invokes the xBestIndex() 1044 ** method of the virtual table with the sqlite3_index_info object that 1045 ** comes in as the 3rd argument to this function. 1046 ** 1047 ** If an error occurs, pParse is populated with an error message and an 1048 ** appropriate error code is returned. A return of SQLITE_CONSTRAINT from 1049 ** xBestIndex is not considered an error. SQLITE_CONSTRAINT indicates that 1050 ** the current configuration of "unusable" flags in sqlite3_index_info can 1051 ** not result in a valid plan. 1052 ** 1053 ** Whether or not an error is returned, it is the responsibility of the 1054 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 1055 ** that this is required. 1056 */ 1057 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 1058 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 1059 int rc; 1060 1061 TRACE_IDX_INPUTS(p); 1062 rc = pVtab->pModule->xBestIndex(pVtab, p); 1063 TRACE_IDX_OUTPUTS(p); 1064 1065 if( rc!=SQLITE_OK && rc!=SQLITE_CONSTRAINT ){ 1066 if( rc==SQLITE_NOMEM ){ 1067 sqlite3OomFault(pParse->db); 1068 }else if( !pVtab->zErrMsg ){ 1069 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 1070 }else{ 1071 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 1072 } 1073 } 1074 sqlite3_free(pVtab->zErrMsg); 1075 pVtab->zErrMsg = 0; 1076 return rc; 1077 } 1078 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 1079 1080 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1081 /* 1082 ** Estimate the location of a particular key among all keys in an 1083 ** index. Store the results in aStat as follows: 1084 ** 1085 ** aStat[0] Est. number of rows less than pRec 1086 ** aStat[1] Est. number of rows equal to pRec 1087 ** 1088 ** Return the index of the sample that is the smallest sample that 1089 ** is greater than or equal to pRec. Note that this index is not an index 1090 ** into the aSample[] array - it is an index into a virtual set of samples 1091 ** based on the contents of aSample[] and the number of fields in record 1092 ** pRec. 1093 */ 1094 static int whereKeyStats( 1095 Parse *pParse, /* Database connection */ 1096 Index *pIdx, /* Index to consider domain of */ 1097 UnpackedRecord *pRec, /* Vector of values to consider */ 1098 int roundUp, /* Round up if true. Round down if false */ 1099 tRowcnt *aStat /* OUT: stats written here */ 1100 ){ 1101 IndexSample *aSample = pIdx->aSample; 1102 int iCol; /* Index of required stats in anEq[] etc. */ 1103 int i; /* Index of first sample >= pRec */ 1104 int iSample; /* Smallest sample larger than or equal to pRec */ 1105 int iMin = 0; /* Smallest sample not yet tested */ 1106 int iTest; /* Next sample to test */ 1107 int res; /* Result of comparison operation */ 1108 int nField; /* Number of fields in pRec */ 1109 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 1110 1111 #ifndef SQLITE_DEBUG 1112 UNUSED_PARAMETER( pParse ); 1113 #endif 1114 assert( pRec!=0 ); 1115 assert( pIdx->nSample>0 ); 1116 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1117 1118 /* Do a binary search to find the first sample greater than or equal 1119 ** to pRec. If pRec contains a single field, the set of samples to search 1120 ** is simply the aSample[] array. If the samples in aSample[] contain more 1121 ** than one fields, all fields following the first are ignored. 1122 ** 1123 ** If pRec contains N fields, where N is more than one, then as well as the 1124 ** samples in aSample[] (truncated to N fields), the search also has to 1125 ** consider prefixes of those samples. For example, if the set of samples 1126 ** in aSample is: 1127 ** 1128 ** aSample[0] = (a, 5) 1129 ** aSample[1] = (a, 10) 1130 ** aSample[2] = (b, 5) 1131 ** aSample[3] = (c, 100) 1132 ** aSample[4] = (c, 105) 1133 ** 1134 ** Then the search space should ideally be the samples above and the 1135 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1136 ** the code actually searches this set: 1137 ** 1138 ** 0: (a) 1139 ** 1: (a, 5) 1140 ** 2: (a, 10) 1141 ** 3: (a, 10) 1142 ** 4: (b) 1143 ** 5: (b, 5) 1144 ** 6: (c) 1145 ** 7: (c, 100) 1146 ** 8: (c, 105) 1147 ** 9: (c, 105) 1148 ** 1149 ** For each sample in the aSample[] array, N samples are present in the 1150 ** effective sample array. In the above, samples 0 and 1 are based on 1151 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1152 ** 1153 ** Often, sample i of each block of N effective samples has (i+1) fields. 1154 ** Except, each sample may be extended to ensure that it is greater than or 1155 ** equal to the previous sample in the array. For example, in the above, 1156 ** sample 2 is the first sample of a block of N samples, so at first it 1157 ** appears that it should be 1 field in size. However, that would make it 1158 ** smaller than sample 1, so the binary search would not work. As a result, 1159 ** it is extended to two fields. The duplicates that this creates do not 1160 ** cause any problems. 1161 */ 1162 nField = pRec->nField; 1163 iCol = 0; 1164 iSample = pIdx->nSample * nField; 1165 do{ 1166 int iSamp; /* Index in aSample[] of test sample */ 1167 int n; /* Number of fields in test sample */ 1168 1169 iTest = (iMin+iSample)/2; 1170 iSamp = iTest / nField; 1171 if( iSamp>0 ){ 1172 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1173 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1174 ** fields that is greater than the previous effective sample. */ 1175 for(n=(iTest % nField) + 1; n<nField; n++){ 1176 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1177 } 1178 }else{ 1179 n = iTest + 1; 1180 } 1181 1182 pRec->nField = n; 1183 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1184 if( res<0 ){ 1185 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1186 iMin = iTest+1; 1187 }else if( res==0 && n<nField ){ 1188 iLower = aSample[iSamp].anLt[n-1]; 1189 iMin = iTest+1; 1190 res = -1; 1191 }else{ 1192 iSample = iTest; 1193 iCol = n-1; 1194 } 1195 }while( res && iMin<iSample ); 1196 i = iSample / nField; 1197 1198 #ifdef SQLITE_DEBUG 1199 /* The following assert statements check that the binary search code 1200 ** above found the right answer. This block serves no purpose other 1201 ** than to invoke the asserts. */ 1202 if( pParse->db->mallocFailed==0 ){ 1203 if( res==0 ){ 1204 /* If (res==0) is true, then pRec must be equal to sample i. */ 1205 assert( i<pIdx->nSample ); 1206 assert( iCol==nField-1 ); 1207 pRec->nField = nField; 1208 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1209 || pParse->db->mallocFailed 1210 ); 1211 }else{ 1212 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1213 ** all samples in the aSample[] array, pRec must be smaller than the 1214 ** (iCol+1) field prefix of sample i. */ 1215 assert( i<=pIdx->nSample && i>=0 ); 1216 pRec->nField = iCol+1; 1217 assert( i==pIdx->nSample 1218 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1219 || pParse->db->mallocFailed ); 1220 1221 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1222 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1223 ** be greater than or equal to the (iCol) field prefix of sample i. 1224 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1225 if( iCol>0 ){ 1226 pRec->nField = iCol; 1227 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1228 || pParse->db->mallocFailed ); 1229 } 1230 if( i>0 ){ 1231 pRec->nField = nField; 1232 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1233 || pParse->db->mallocFailed ); 1234 } 1235 } 1236 } 1237 #endif /* ifdef SQLITE_DEBUG */ 1238 1239 if( res==0 ){ 1240 /* Record pRec is equal to sample i */ 1241 assert( iCol==nField-1 ); 1242 aStat[0] = aSample[i].anLt[iCol]; 1243 aStat[1] = aSample[i].anEq[iCol]; 1244 }else{ 1245 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1246 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1247 ** is larger than all samples in the array. */ 1248 tRowcnt iUpper, iGap; 1249 if( i>=pIdx->nSample ){ 1250 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1251 }else{ 1252 iUpper = aSample[i].anLt[iCol]; 1253 } 1254 1255 if( iLower>=iUpper ){ 1256 iGap = 0; 1257 }else{ 1258 iGap = iUpper - iLower; 1259 } 1260 if( roundUp ){ 1261 iGap = (iGap*2)/3; 1262 }else{ 1263 iGap = iGap/3; 1264 } 1265 aStat[0] = iLower + iGap; 1266 aStat[1] = pIdx->aAvgEq[nField-1]; 1267 } 1268 1269 /* Restore the pRec->nField value before returning. */ 1270 pRec->nField = nField; 1271 return i; 1272 } 1273 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1274 1275 /* 1276 ** If it is not NULL, pTerm is a term that provides an upper or lower 1277 ** bound on a range scan. Without considering pTerm, it is estimated 1278 ** that the scan will visit nNew rows. This function returns the number 1279 ** estimated to be visited after taking pTerm into account. 1280 ** 1281 ** If the user explicitly specified a likelihood() value for this term, 1282 ** then the return value is the likelihood multiplied by the number of 1283 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1284 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1285 */ 1286 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1287 LogEst nRet = nNew; 1288 if( pTerm ){ 1289 if( pTerm->truthProb<=0 ){ 1290 nRet += pTerm->truthProb; 1291 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1292 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1293 } 1294 } 1295 return nRet; 1296 } 1297 1298 1299 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1300 /* 1301 ** Return the affinity for a single column of an index. 1302 */ 1303 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1304 assert( iCol>=0 && iCol<pIdx->nColumn ); 1305 if( !pIdx->zColAff ){ 1306 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1307 } 1308 return pIdx->zColAff[iCol]; 1309 } 1310 #endif 1311 1312 1313 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1314 /* 1315 ** This function is called to estimate the number of rows visited by a 1316 ** range-scan on a skip-scan index. For example: 1317 ** 1318 ** CREATE INDEX i1 ON t1(a, b, c); 1319 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1320 ** 1321 ** Value pLoop->nOut is currently set to the estimated number of rows 1322 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1323 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1324 ** on the stat4 data for the index. this scan will be peformed multiple 1325 ** times (once for each (a,b) combination that matches a=?) is dealt with 1326 ** by the caller. 1327 ** 1328 ** It does this by scanning through all stat4 samples, comparing values 1329 ** extracted from pLower and pUpper with the corresponding column in each 1330 ** sample. If L and U are the number of samples found to be less than or 1331 ** equal to the values extracted from pLower and pUpper respectively, and 1332 ** N is the total number of samples, the pLoop->nOut value is adjusted 1333 ** as follows: 1334 ** 1335 ** nOut = nOut * ( min(U - L, 1) / N ) 1336 ** 1337 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1338 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1339 ** U is set to N. 1340 ** 1341 ** Normally, this function sets *pbDone to 1 before returning. However, 1342 ** if no value can be extracted from either pLower or pUpper (and so the 1343 ** estimate of the number of rows delivered remains unchanged), *pbDone 1344 ** is left as is. 1345 ** 1346 ** If an error occurs, an SQLite error code is returned. Otherwise, 1347 ** SQLITE_OK. 1348 */ 1349 static int whereRangeSkipScanEst( 1350 Parse *pParse, /* Parsing & code generating context */ 1351 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1352 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1353 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1354 int *pbDone /* Set to true if at least one expr. value extracted */ 1355 ){ 1356 Index *p = pLoop->u.btree.pIndex; 1357 int nEq = pLoop->u.btree.nEq; 1358 sqlite3 *db = pParse->db; 1359 int nLower = -1; 1360 int nUpper = p->nSample+1; 1361 int rc = SQLITE_OK; 1362 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1363 CollSeq *pColl; 1364 1365 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1366 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1367 sqlite3_value *pVal = 0; /* Value extracted from record */ 1368 1369 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1370 if( pLower ){ 1371 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1372 nLower = 0; 1373 } 1374 if( pUpper && rc==SQLITE_OK ){ 1375 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1376 nUpper = p2 ? 0 : p->nSample; 1377 } 1378 1379 if( p1 || p2 ){ 1380 int i; 1381 int nDiff; 1382 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1383 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1384 if( rc==SQLITE_OK && p1 ){ 1385 int res = sqlite3MemCompare(p1, pVal, pColl); 1386 if( res>=0 ) nLower++; 1387 } 1388 if( rc==SQLITE_OK && p2 ){ 1389 int res = sqlite3MemCompare(p2, pVal, pColl); 1390 if( res>=0 ) nUpper++; 1391 } 1392 } 1393 nDiff = (nUpper - nLower); 1394 if( nDiff<=0 ) nDiff = 1; 1395 1396 /* If there is both an upper and lower bound specified, and the 1397 ** comparisons indicate that they are close together, use the fallback 1398 ** method (assume that the scan visits 1/64 of the rows) for estimating 1399 ** the number of rows visited. Otherwise, estimate the number of rows 1400 ** using the method described in the header comment for this function. */ 1401 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1402 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1403 pLoop->nOut -= nAdjust; 1404 *pbDone = 1; 1405 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1406 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1407 } 1408 1409 }else{ 1410 assert( *pbDone==0 ); 1411 } 1412 1413 sqlite3ValueFree(p1); 1414 sqlite3ValueFree(p2); 1415 sqlite3ValueFree(pVal); 1416 1417 return rc; 1418 } 1419 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1420 1421 /* 1422 ** This function is used to estimate the number of rows that will be visited 1423 ** by scanning an index for a range of values. The range may have an upper 1424 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1425 ** and lower bounds are represented by pLower and pUpper respectively. For 1426 ** example, assuming that index p is on t1(a): 1427 ** 1428 ** ... FROM t1 WHERE a > ? AND a < ? ... 1429 ** |_____| |_____| 1430 ** | | 1431 ** pLower pUpper 1432 ** 1433 ** If either of the upper or lower bound is not present, then NULL is passed in 1434 ** place of the corresponding WhereTerm. 1435 ** 1436 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1437 ** column subject to the range constraint. Or, equivalently, the number of 1438 ** equality constraints optimized by the proposed index scan. For example, 1439 ** assuming index p is on t1(a, b), and the SQL query is: 1440 ** 1441 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1442 ** 1443 ** then nEq is set to 1 (as the range restricted column, b, is the second 1444 ** left-most column of the index). Or, if the query is: 1445 ** 1446 ** ... FROM t1 WHERE a > ? AND a < ? ... 1447 ** 1448 ** then nEq is set to 0. 1449 ** 1450 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1451 ** number of rows that the index scan is expected to visit without 1452 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1453 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1454 ** to account for the range constraints pLower and pUpper. 1455 ** 1456 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1457 ** used, a single range inequality reduces the search space by a factor of 4. 1458 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1459 ** rows visited by a factor of 64. 1460 */ 1461 static int whereRangeScanEst( 1462 Parse *pParse, /* Parsing & code generating context */ 1463 WhereLoopBuilder *pBuilder, 1464 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1465 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1466 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1467 ){ 1468 int rc = SQLITE_OK; 1469 int nOut = pLoop->nOut; 1470 LogEst nNew; 1471 1472 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1473 Index *p = pLoop->u.btree.pIndex; 1474 int nEq = pLoop->u.btree.nEq; 1475 1476 if( p->nSample>0 && nEq<p->nSampleCol 1477 && OptimizationEnabled(pParse->db, SQLITE_Stat34) 1478 ){ 1479 if( nEq==pBuilder->nRecValid ){ 1480 UnpackedRecord *pRec = pBuilder->pRec; 1481 tRowcnt a[2]; 1482 int nBtm = pLoop->u.btree.nBtm; 1483 int nTop = pLoop->u.btree.nTop; 1484 1485 /* Variable iLower will be set to the estimate of the number of rows in 1486 ** the index that are less than the lower bound of the range query. The 1487 ** lower bound being the concatenation of $P and $L, where $P is the 1488 ** key-prefix formed by the nEq values matched against the nEq left-most 1489 ** columns of the index, and $L is the value in pLower. 1490 ** 1491 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1492 ** is not a simple variable or literal value), the lower bound of the 1493 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1494 ** if $L is available, whereKeyStats() is called for both ($P) and 1495 ** ($P:$L) and the larger of the two returned values is used. 1496 ** 1497 ** Similarly, iUpper is to be set to the estimate of the number of rows 1498 ** less than the upper bound of the range query. Where the upper bound 1499 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1500 ** of iUpper are requested of whereKeyStats() and the smaller used. 1501 ** 1502 ** The number of rows between the two bounds is then just iUpper-iLower. 1503 */ 1504 tRowcnt iLower; /* Rows less than the lower bound */ 1505 tRowcnt iUpper; /* Rows less than the upper bound */ 1506 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1507 int iUprIdx = -1; /* aSample[] for the upper bound */ 1508 1509 if( pRec ){ 1510 testcase( pRec->nField!=pBuilder->nRecValid ); 1511 pRec->nField = pBuilder->nRecValid; 1512 } 1513 /* Determine iLower and iUpper using ($P) only. */ 1514 if( nEq==0 ){ 1515 iLower = 0; 1516 iUpper = p->nRowEst0; 1517 }else{ 1518 /* Note: this call could be optimized away - since the same values must 1519 ** have been requested when testing key $P in whereEqualScanEst(). */ 1520 whereKeyStats(pParse, p, pRec, 0, a); 1521 iLower = a[0]; 1522 iUpper = a[0] + a[1]; 1523 } 1524 1525 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1526 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1527 assert( p->aSortOrder!=0 ); 1528 if( p->aSortOrder[nEq] ){ 1529 /* The roles of pLower and pUpper are swapped for a DESC index */ 1530 SWAP(WhereTerm*, pLower, pUpper); 1531 SWAP(int, nBtm, nTop); 1532 } 1533 1534 /* If possible, improve on the iLower estimate using ($P:$L). */ 1535 if( pLower ){ 1536 int n; /* Values extracted from pExpr */ 1537 Expr *pExpr = pLower->pExpr->pRight; 1538 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1539 if( rc==SQLITE_OK && n ){ 1540 tRowcnt iNew; 1541 u16 mask = WO_GT|WO_LE; 1542 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1543 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1544 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1545 if( iNew>iLower ) iLower = iNew; 1546 nOut--; 1547 pLower = 0; 1548 } 1549 } 1550 1551 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1552 if( pUpper ){ 1553 int n; /* Values extracted from pExpr */ 1554 Expr *pExpr = pUpper->pExpr->pRight; 1555 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1556 if( rc==SQLITE_OK && n ){ 1557 tRowcnt iNew; 1558 u16 mask = WO_GT|WO_LE; 1559 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1560 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1561 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1562 if( iNew<iUpper ) iUpper = iNew; 1563 nOut--; 1564 pUpper = 0; 1565 } 1566 } 1567 1568 pBuilder->pRec = pRec; 1569 if( rc==SQLITE_OK ){ 1570 if( iUpper>iLower ){ 1571 nNew = sqlite3LogEst(iUpper - iLower); 1572 /* TUNING: If both iUpper and iLower are derived from the same 1573 ** sample, then assume they are 4x more selective. This brings 1574 ** the estimated selectivity more in line with what it would be 1575 ** if estimated without the use of STAT3/4 tables. */ 1576 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1577 }else{ 1578 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1579 } 1580 if( nNew<nOut ){ 1581 nOut = nNew; 1582 } 1583 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1584 (u32)iLower, (u32)iUpper, nOut)); 1585 } 1586 }else{ 1587 int bDone = 0; 1588 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1589 if( bDone ) return rc; 1590 } 1591 } 1592 #else 1593 UNUSED_PARAMETER(pParse); 1594 UNUSED_PARAMETER(pBuilder); 1595 assert( pLower || pUpper ); 1596 #endif 1597 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1598 nNew = whereRangeAdjust(pLower, nOut); 1599 nNew = whereRangeAdjust(pUpper, nNew); 1600 1601 /* TUNING: If there is both an upper and lower limit and neither limit 1602 ** has an application-defined likelihood(), assume the range is 1603 ** reduced by an additional 75%. This means that, by default, an open-ended 1604 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1605 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1606 ** match 1/64 of the index. */ 1607 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1608 nNew -= 20; 1609 } 1610 1611 nOut -= (pLower!=0) + (pUpper!=0); 1612 if( nNew<10 ) nNew = 10; 1613 if( nNew<nOut ) nOut = nNew; 1614 #if defined(WHERETRACE_ENABLED) 1615 if( pLoop->nOut>nOut ){ 1616 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1617 pLoop->nOut, nOut)); 1618 } 1619 #endif 1620 pLoop->nOut = (LogEst)nOut; 1621 return rc; 1622 } 1623 1624 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1625 /* 1626 ** Estimate the number of rows that will be returned based on 1627 ** an equality constraint x=VALUE and where that VALUE occurs in 1628 ** the histogram data. This only works when x is the left-most 1629 ** column of an index and sqlite_stat3 histogram data is available 1630 ** for that index. When pExpr==NULL that means the constraint is 1631 ** "x IS NULL" instead of "x=VALUE". 1632 ** 1633 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1634 ** If unable to make an estimate, leave *pnRow unchanged and return 1635 ** non-zero. 1636 ** 1637 ** This routine can fail if it is unable to load a collating sequence 1638 ** required for string comparison, or if unable to allocate memory 1639 ** for a UTF conversion required for comparison. The error is stored 1640 ** in the pParse structure. 1641 */ 1642 static int whereEqualScanEst( 1643 Parse *pParse, /* Parsing & code generating context */ 1644 WhereLoopBuilder *pBuilder, 1645 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1646 tRowcnt *pnRow /* Write the revised row estimate here */ 1647 ){ 1648 Index *p = pBuilder->pNew->u.btree.pIndex; 1649 int nEq = pBuilder->pNew->u.btree.nEq; 1650 UnpackedRecord *pRec = pBuilder->pRec; 1651 int rc; /* Subfunction return code */ 1652 tRowcnt a[2]; /* Statistics */ 1653 int bOk; 1654 1655 assert( nEq>=1 ); 1656 assert( nEq<=p->nColumn ); 1657 assert( p->aSample!=0 ); 1658 assert( p->nSample>0 ); 1659 assert( pBuilder->nRecValid<nEq ); 1660 1661 /* If values are not available for all fields of the index to the left 1662 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1663 if( pBuilder->nRecValid<(nEq-1) ){ 1664 return SQLITE_NOTFOUND; 1665 } 1666 1667 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1668 ** below would return the same value. */ 1669 if( nEq>=p->nColumn ){ 1670 *pnRow = 1; 1671 return SQLITE_OK; 1672 } 1673 1674 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1675 pBuilder->pRec = pRec; 1676 if( rc!=SQLITE_OK ) return rc; 1677 if( bOk==0 ) return SQLITE_NOTFOUND; 1678 pBuilder->nRecValid = nEq; 1679 1680 whereKeyStats(pParse, p, pRec, 0, a); 1681 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1682 p->zName, nEq-1, (int)a[1])); 1683 *pnRow = a[1]; 1684 1685 return rc; 1686 } 1687 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1688 1689 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1690 /* 1691 ** Estimate the number of rows that will be returned based on 1692 ** an IN constraint where the right-hand side of the IN operator 1693 ** is a list of values. Example: 1694 ** 1695 ** WHERE x IN (1,2,3,4) 1696 ** 1697 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1698 ** If unable to make an estimate, leave *pnRow unchanged and return 1699 ** non-zero. 1700 ** 1701 ** This routine can fail if it is unable to load a collating sequence 1702 ** required for string comparison, or if unable to allocate memory 1703 ** for a UTF conversion required for comparison. The error is stored 1704 ** in the pParse structure. 1705 */ 1706 static int whereInScanEst( 1707 Parse *pParse, /* Parsing & code generating context */ 1708 WhereLoopBuilder *pBuilder, 1709 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1710 tRowcnt *pnRow /* Write the revised row estimate here */ 1711 ){ 1712 Index *p = pBuilder->pNew->u.btree.pIndex; 1713 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1714 int nRecValid = pBuilder->nRecValid; 1715 int rc = SQLITE_OK; /* Subfunction return code */ 1716 tRowcnt nEst; /* Number of rows for a single term */ 1717 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1718 int i; /* Loop counter */ 1719 1720 assert( p->aSample!=0 ); 1721 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1722 nEst = nRow0; 1723 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1724 nRowEst += nEst; 1725 pBuilder->nRecValid = nRecValid; 1726 } 1727 1728 if( rc==SQLITE_OK ){ 1729 if( nRowEst > nRow0 ) nRowEst = nRow0; 1730 *pnRow = nRowEst; 1731 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1732 } 1733 assert( pBuilder->nRecValid==nRecValid ); 1734 return rc; 1735 } 1736 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1737 1738 1739 #ifdef WHERETRACE_ENABLED 1740 /* 1741 ** Print the content of a WhereTerm object 1742 */ 1743 static void whereTermPrint(WhereTerm *pTerm, int iTerm){ 1744 if( pTerm==0 ){ 1745 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1746 }else{ 1747 char zType[4]; 1748 char zLeft[50]; 1749 memcpy(zType, "...", 4); 1750 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1751 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1752 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1753 if( pTerm->eOperator & WO_SINGLE ){ 1754 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 1755 pTerm->leftCursor, pTerm->u.leftColumn); 1756 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 1757 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%lld", 1758 pTerm->u.pOrInfo->indexable); 1759 }else{ 1760 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 1761 } 1762 sqlite3DebugPrintf( 1763 "TERM-%-3d %p %s %-12s prob=%-3d op=0x%03x wtFlags=0x%04x", 1764 iTerm, pTerm, zType, zLeft, pTerm->truthProb, 1765 pTerm->eOperator, pTerm->wtFlags); 1766 if( pTerm->iField ){ 1767 sqlite3DebugPrintf(" iField=%d\n", pTerm->iField); 1768 }else{ 1769 sqlite3DebugPrintf("\n"); 1770 } 1771 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 1772 } 1773 } 1774 #endif 1775 1776 #ifdef WHERETRACE_ENABLED 1777 /* 1778 ** Show the complete content of a WhereClause 1779 */ 1780 void sqlite3WhereClausePrint(WhereClause *pWC){ 1781 int i; 1782 for(i=0; i<pWC->nTerm; i++){ 1783 whereTermPrint(&pWC->a[i], i); 1784 } 1785 } 1786 #endif 1787 1788 #ifdef WHERETRACE_ENABLED 1789 /* 1790 ** Print a WhereLoop object for debugging purposes 1791 */ 1792 static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){ 1793 WhereInfo *pWInfo = pWC->pWInfo; 1794 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 1795 struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab; 1796 Table *pTab = pItem->pTab; 1797 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 1798 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 1799 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 1800 sqlite3DebugPrintf(" %12s", 1801 pItem->zAlias ? pItem->zAlias : pTab->zName); 1802 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 1803 const char *zName; 1804 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 1805 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 1806 int i = sqlite3Strlen30(zName) - 1; 1807 while( zName[i]!='_' ) i--; 1808 zName += i; 1809 } 1810 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 1811 }else{ 1812 sqlite3DebugPrintf("%20s",""); 1813 } 1814 }else{ 1815 char *z; 1816 if( p->u.vtab.idxStr ){ 1817 z = sqlite3_mprintf("(%d,\"%s\",%x)", 1818 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 1819 }else{ 1820 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 1821 } 1822 sqlite3DebugPrintf(" %-19s", z); 1823 sqlite3_free(z); 1824 } 1825 if( p->wsFlags & WHERE_SKIPSCAN ){ 1826 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 1827 }else{ 1828 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); 1829 } 1830 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 1831 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 1832 int i; 1833 for(i=0; i<p->nLTerm; i++){ 1834 whereTermPrint(p->aLTerm[i], i); 1835 } 1836 } 1837 } 1838 #endif 1839 1840 /* 1841 ** Convert bulk memory into a valid WhereLoop that can be passed 1842 ** to whereLoopClear harmlessly. 1843 */ 1844 static void whereLoopInit(WhereLoop *p){ 1845 p->aLTerm = p->aLTermSpace; 1846 p->nLTerm = 0; 1847 p->nLSlot = ArraySize(p->aLTermSpace); 1848 p->wsFlags = 0; 1849 } 1850 1851 /* 1852 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 1853 */ 1854 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 1855 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 1856 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 1857 sqlite3_free(p->u.vtab.idxStr); 1858 p->u.vtab.needFree = 0; 1859 p->u.vtab.idxStr = 0; 1860 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 1861 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 1862 sqlite3DbFreeNN(db, p->u.btree.pIndex); 1863 p->u.btree.pIndex = 0; 1864 } 1865 } 1866 } 1867 1868 /* 1869 ** Deallocate internal memory used by a WhereLoop object 1870 */ 1871 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 1872 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1873 whereLoopClearUnion(db, p); 1874 whereLoopInit(p); 1875 } 1876 1877 /* 1878 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 1879 */ 1880 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 1881 WhereTerm **paNew; 1882 if( p->nLSlot>=n ) return SQLITE_OK; 1883 n = (n+7)&~7; 1884 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 1885 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 1886 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 1887 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1888 p->aLTerm = paNew; 1889 p->nLSlot = n; 1890 return SQLITE_OK; 1891 } 1892 1893 /* 1894 ** Transfer content from the second pLoop into the first. 1895 */ 1896 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 1897 whereLoopClearUnion(db, pTo); 1898 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 1899 memset(&pTo->u, 0, sizeof(pTo->u)); 1900 return SQLITE_NOMEM_BKPT; 1901 } 1902 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 1903 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 1904 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 1905 pFrom->u.vtab.needFree = 0; 1906 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 1907 pFrom->u.btree.pIndex = 0; 1908 } 1909 return SQLITE_OK; 1910 } 1911 1912 /* 1913 ** Delete a WhereLoop object 1914 */ 1915 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 1916 whereLoopClear(db, p); 1917 sqlite3DbFreeNN(db, p); 1918 } 1919 1920 /* 1921 ** Free a WhereInfo structure 1922 */ 1923 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 1924 int i; 1925 assert( pWInfo!=0 ); 1926 for(i=0; i<pWInfo->nLevel; i++){ 1927 WhereLevel *pLevel = &pWInfo->a[i]; 1928 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){ 1929 sqlite3DbFree(db, pLevel->u.in.aInLoop); 1930 } 1931 } 1932 sqlite3WhereClauseClear(&pWInfo->sWC); 1933 while( pWInfo->pLoops ){ 1934 WhereLoop *p = pWInfo->pLoops; 1935 pWInfo->pLoops = p->pNextLoop; 1936 whereLoopDelete(db, p); 1937 } 1938 sqlite3DbFreeNN(db, pWInfo); 1939 } 1940 1941 /* 1942 ** Return TRUE if all of the following are true: 1943 ** 1944 ** (1) X has the same or lower cost that Y 1945 ** (2) X uses fewer WHERE clause terms than Y 1946 ** (3) Every WHERE clause term used by X is also used by Y 1947 ** (4) X skips at least as many columns as Y 1948 ** (5) If X is a covering index, than Y is too 1949 ** 1950 ** Conditions (2) and (3) mean that X is a "proper subset" of Y. 1951 ** If X is a proper subset of Y then Y is a better choice and ought 1952 ** to have a lower cost. This routine returns TRUE when that cost 1953 ** relationship is inverted and needs to be adjusted. Constraint (4) 1954 ** was added because if X uses skip-scan less than Y it still might 1955 ** deserve a lower cost even if it is a proper subset of Y. Constraint (5) 1956 ** was added because a covering index probably deserves to have a lower cost 1957 ** than a non-covering index even if it is a proper subset. 1958 */ 1959 static int whereLoopCheaperProperSubset( 1960 const WhereLoop *pX, /* First WhereLoop to compare */ 1961 const WhereLoop *pY /* Compare against this WhereLoop */ 1962 ){ 1963 int i, j; 1964 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 1965 return 0; /* X is not a subset of Y */ 1966 } 1967 if( pY->nSkip > pX->nSkip ) return 0; 1968 if( pX->rRun >= pY->rRun ){ 1969 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ 1970 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ 1971 } 1972 for(i=pX->nLTerm-1; i>=0; i--){ 1973 if( pX->aLTerm[i]==0 ) continue; 1974 for(j=pY->nLTerm-1; j>=0; j--){ 1975 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 1976 } 1977 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 1978 } 1979 if( (pX->wsFlags&WHERE_IDX_ONLY)!=0 1980 && (pY->wsFlags&WHERE_IDX_ONLY)==0 ){ 1981 return 0; /* Constraint (5) */ 1982 } 1983 return 1; /* All conditions meet */ 1984 } 1985 1986 /* 1987 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so 1988 ** that: 1989 ** 1990 ** (1) pTemplate costs less than any other WhereLoops that are a proper 1991 ** subset of pTemplate 1992 ** 1993 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 1994 ** is a proper subset. 1995 ** 1996 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 1997 ** WHERE clause terms than Y and that every WHERE clause term used by X is 1998 ** also used by Y. 1999 */ 2000 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 2001 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 2002 for(; p; p=p->pNextLoop){ 2003 if( p->iTab!=pTemplate->iTab ) continue; 2004 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 2005 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 2006 /* Adjust pTemplate cost downward so that it is cheaper than its 2007 ** subset p. */ 2008 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2009 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); 2010 pTemplate->rRun = p->rRun; 2011 pTemplate->nOut = p->nOut - 1; 2012 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 2013 /* Adjust pTemplate cost upward so that it is costlier than p since 2014 ** pTemplate is a proper subset of p */ 2015 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2016 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); 2017 pTemplate->rRun = p->rRun; 2018 pTemplate->nOut = p->nOut + 1; 2019 } 2020 } 2021 } 2022 2023 /* 2024 ** Search the list of WhereLoops in *ppPrev looking for one that can be 2025 ** replaced by pTemplate. 2026 ** 2027 ** Return NULL if pTemplate does not belong on the WhereLoop list. 2028 ** In other words if pTemplate ought to be dropped from further consideration. 2029 ** 2030 ** If pX is a WhereLoop that pTemplate can replace, then return the 2031 ** link that points to pX. 2032 ** 2033 ** If pTemplate cannot replace any existing element of the list but needs 2034 ** to be added to the list as a new entry, then return a pointer to the 2035 ** tail of the list. 2036 */ 2037 static WhereLoop **whereLoopFindLesser( 2038 WhereLoop **ppPrev, 2039 const WhereLoop *pTemplate 2040 ){ 2041 WhereLoop *p; 2042 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 2043 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 2044 /* If either the iTab or iSortIdx values for two WhereLoop are different 2045 ** then those WhereLoops need to be considered separately. Neither is 2046 ** a candidate to replace the other. */ 2047 continue; 2048 } 2049 /* In the current implementation, the rSetup value is either zero 2050 ** or the cost of building an automatic index (NlogN) and the NlogN 2051 ** is the same for compatible WhereLoops. */ 2052 assert( p->rSetup==0 || pTemplate->rSetup==0 2053 || p->rSetup==pTemplate->rSetup ); 2054 2055 /* whereLoopAddBtree() always generates and inserts the automatic index 2056 ** case first. Hence compatible candidate WhereLoops never have a larger 2057 ** rSetup. Call this SETUP-INVARIANT */ 2058 assert( p->rSetup>=pTemplate->rSetup ); 2059 2060 /* Any loop using an appliation-defined index (or PRIMARY KEY or 2061 ** UNIQUE constraint) with one or more == constraints is better 2062 ** than an automatic index. Unless it is a skip-scan. */ 2063 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 2064 && (pTemplate->nSkip)==0 2065 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 2066 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 2067 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2068 ){ 2069 break; 2070 } 2071 2072 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2073 ** discarded. WhereLoop p is better if: 2074 ** (1) p has no more dependencies than pTemplate, and 2075 ** (2) p has an equal or lower cost than pTemplate 2076 */ 2077 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2078 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2079 && p->rRun<=pTemplate->rRun /* (2b) */ 2080 && p->nOut<=pTemplate->nOut /* (2c) */ 2081 ){ 2082 return 0; /* Discard pTemplate */ 2083 } 2084 2085 /* If pTemplate is always better than p, then cause p to be overwritten 2086 ** with pTemplate. pTemplate is better than p if: 2087 ** (1) pTemplate has no more dependences than p, and 2088 ** (2) pTemplate has an equal or lower cost than p. 2089 */ 2090 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2091 && p->rRun>=pTemplate->rRun /* (2a) */ 2092 && p->nOut>=pTemplate->nOut /* (2b) */ 2093 ){ 2094 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2095 break; /* Cause p to be overwritten by pTemplate */ 2096 } 2097 } 2098 return ppPrev; 2099 } 2100 2101 /* 2102 ** Insert or replace a WhereLoop entry using the template supplied. 2103 ** 2104 ** An existing WhereLoop entry might be overwritten if the new template 2105 ** is better and has fewer dependencies. Or the template will be ignored 2106 ** and no insert will occur if an existing WhereLoop is faster and has 2107 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2108 ** added based on the template. 2109 ** 2110 ** If pBuilder->pOrSet is not NULL then we care about only the 2111 ** prerequisites and rRun and nOut costs of the N best loops. That 2112 ** information is gathered in the pBuilder->pOrSet object. This special 2113 ** processing mode is used only for OR clause processing. 2114 ** 2115 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2116 ** still might overwrite similar loops with the new template if the 2117 ** new template is better. Loops may be overwritten if the following 2118 ** conditions are met: 2119 ** 2120 ** (1) They have the same iTab. 2121 ** (2) They have the same iSortIdx. 2122 ** (3) The template has same or fewer dependencies than the current loop 2123 ** (4) The template has the same or lower cost than the current loop 2124 */ 2125 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2126 WhereLoop **ppPrev, *p; 2127 WhereInfo *pWInfo = pBuilder->pWInfo; 2128 sqlite3 *db = pWInfo->pParse->db; 2129 int rc; 2130 2131 /* Stop the search once we hit the query planner search limit */ 2132 if( pBuilder->iPlanLimit==0 ){ 2133 WHERETRACE(0xffffffff,("=== query planner search limit reached ===\n")); 2134 if( pBuilder->pOrSet ) pBuilder->pOrSet->n = 0; 2135 return SQLITE_DONE; 2136 } 2137 pBuilder->iPlanLimit--; 2138 2139 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2140 ** and prereqs. 2141 */ 2142 if( pBuilder->pOrSet!=0 ){ 2143 if( pTemplate->nLTerm ){ 2144 #if WHERETRACE_ENABLED 2145 u16 n = pBuilder->pOrSet->n; 2146 int x = 2147 #endif 2148 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2149 pTemplate->nOut); 2150 #if WHERETRACE_ENABLED /* 0x8 */ 2151 if( sqlite3WhereTrace & 0x8 ){ 2152 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2153 whereLoopPrint(pTemplate, pBuilder->pWC); 2154 } 2155 #endif 2156 } 2157 return SQLITE_OK; 2158 } 2159 2160 /* Look for an existing WhereLoop to replace with pTemplate 2161 */ 2162 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2163 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2164 2165 if( ppPrev==0 ){ 2166 /* There already exists a WhereLoop on the list that is better 2167 ** than pTemplate, so just ignore pTemplate */ 2168 #if WHERETRACE_ENABLED /* 0x8 */ 2169 if( sqlite3WhereTrace & 0x8 ){ 2170 sqlite3DebugPrintf(" skip: "); 2171 whereLoopPrint(pTemplate, pBuilder->pWC); 2172 } 2173 #endif 2174 return SQLITE_OK; 2175 }else{ 2176 p = *ppPrev; 2177 } 2178 2179 /* If we reach this point it means that either p[] should be overwritten 2180 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2181 ** WhereLoop and insert it. 2182 */ 2183 #if WHERETRACE_ENABLED /* 0x8 */ 2184 if( sqlite3WhereTrace & 0x8 ){ 2185 if( p!=0 ){ 2186 sqlite3DebugPrintf("replace: "); 2187 whereLoopPrint(p, pBuilder->pWC); 2188 sqlite3DebugPrintf(" with: "); 2189 }else{ 2190 sqlite3DebugPrintf(" add: "); 2191 } 2192 whereLoopPrint(pTemplate, pBuilder->pWC); 2193 } 2194 #endif 2195 if( p==0 ){ 2196 /* Allocate a new WhereLoop to add to the end of the list */ 2197 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2198 if( p==0 ) return SQLITE_NOMEM_BKPT; 2199 whereLoopInit(p); 2200 p->pNextLoop = 0; 2201 }else{ 2202 /* We will be overwriting WhereLoop p[]. But before we do, first 2203 ** go through the rest of the list and delete any other entries besides 2204 ** p[] that are also supplated by pTemplate */ 2205 WhereLoop **ppTail = &p->pNextLoop; 2206 WhereLoop *pToDel; 2207 while( *ppTail ){ 2208 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2209 if( ppTail==0 ) break; 2210 pToDel = *ppTail; 2211 if( pToDel==0 ) break; 2212 *ppTail = pToDel->pNextLoop; 2213 #if WHERETRACE_ENABLED /* 0x8 */ 2214 if( sqlite3WhereTrace & 0x8 ){ 2215 sqlite3DebugPrintf(" delete: "); 2216 whereLoopPrint(pToDel, pBuilder->pWC); 2217 } 2218 #endif 2219 whereLoopDelete(db, pToDel); 2220 } 2221 } 2222 rc = whereLoopXfer(db, p, pTemplate); 2223 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2224 Index *pIndex = p->u.btree.pIndex; 2225 if( pIndex && pIndex->idxType==SQLITE_IDXTYPE_IPK ){ 2226 p->u.btree.pIndex = 0; 2227 } 2228 } 2229 return rc; 2230 } 2231 2232 /* 2233 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2234 ** WHERE clause that reference the loop but which are not used by an 2235 ** index. 2236 * 2237 ** For every WHERE clause term that is not used by the index 2238 ** and which has a truth probability assigned by one of the likelihood(), 2239 ** likely(), or unlikely() SQL functions, reduce the estimated number 2240 ** of output rows by the probability specified. 2241 ** 2242 ** TUNING: For every WHERE clause term that is not used by the index 2243 ** and which does not have an assigned truth probability, heuristics 2244 ** described below are used to try to estimate the truth probability. 2245 ** TODO --> Perhaps this is something that could be improved by better 2246 ** table statistics. 2247 ** 2248 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2249 ** value corresponds to -1 in LogEst notation, so this means decrement 2250 ** the WhereLoop.nOut field for every such WHERE clause term. 2251 ** 2252 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2253 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2254 ** final output row estimate is no greater than 1/4 of the total number 2255 ** of rows in the table. In other words, assume that x==EXPR will filter 2256 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2257 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2258 ** on the "x" column and so in that case only cap the output row estimate 2259 ** at 1/2 instead of 1/4. 2260 */ 2261 static void whereLoopOutputAdjust( 2262 WhereClause *pWC, /* The WHERE clause */ 2263 WhereLoop *pLoop, /* The loop to adjust downward */ 2264 LogEst nRow /* Number of rows in the entire table */ 2265 ){ 2266 WhereTerm *pTerm, *pX; 2267 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2268 int i, j, k; 2269 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2270 2271 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2272 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ 2273 assert( pTerm!=0 ); 2274 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; 2275 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2276 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2277 for(j=pLoop->nLTerm-1; j>=0; j--){ 2278 pX = pLoop->aLTerm[j]; 2279 if( pX==0 ) continue; 2280 if( pX==pTerm ) break; 2281 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2282 } 2283 if( j<0 ){ 2284 if( pTerm->truthProb<=0 ){ 2285 /* If a truth probability is specified using the likelihood() hints, 2286 ** then use the probability provided by the application. */ 2287 pLoop->nOut += pTerm->truthProb; 2288 }else{ 2289 /* In the absence of explicit truth probabilities, use heuristics to 2290 ** guess a reasonable truth probability. */ 2291 pLoop->nOut--; 2292 if( pTerm->eOperator&(WO_EQ|WO_IS) ){ 2293 Expr *pRight = pTerm->pExpr->pRight; 2294 testcase( pTerm->pExpr->op==TK_IS ); 2295 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2296 k = 10; 2297 }else{ 2298 k = 20; 2299 } 2300 if( iReduce<k ) iReduce = k; 2301 } 2302 } 2303 } 2304 } 2305 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce; 2306 } 2307 2308 /* 2309 ** Term pTerm is a vector range comparison operation. The first comparison 2310 ** in the vector can be optimized using column nEq of the index. This 2311 ** function returns the total number of vector elements that can be used 2312 ** as part of the range comparison. 2313 ** 2314 ** For example, if the query is: 2315 ** 2316 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2317 ** 2318 ** and the index: 2319 ** 2320 ** CREATE INDEX ... ON (a, b, c, d, e) 2321 ** 2322 ** then this function would be invoked with nEq=1. The value returned in 2323 ** this case is 3. 2324 */ 2325 static int whereRangeVectorLen( 2326 Parse *pParse, /* Parsing context */ 2327 int iCur, /* Cursor open on pIdx */ 2328 Index *pIdx, /* The index to be used for a inequality constraint */ 2329 int nEq, /* Number of prior equality constraints on same index */ 2330 WhereTerm *pTerm /* The vector inequality constraint */ 2331 ){ 2332 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2333 int i; 2334 2335 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2336 for(i=1; i<nCmp; i++){ 2337 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2338 ** of the index. If not, exit the loop. */ 2339 char aff; /* Comparison affinity */ 2340 char idxaff = 0; /* Indexed columns affinity */ 2341 CollSeq *pColl; /* Comparison collation sequence */ 2342 Expr *pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2343 Expr *pRhs = pTerm->pExpr->pRight; 2344 if( pRhs->flags & EP_xIsSelect ){ 2345 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2346 }else{ 2347 pRhs = pRhs->x.pList->a[i].pExpr; 2348 } 2349 2350 /* Check that the LHS of the comparison is a column reference to 2351 ** the right column of the right source table. And that the sort 2352 ** order of the index column is the same as the sort order of the 2353 ** leftmost index column. */ 2354 if( pLhs->op!=TK_COLUMN 2355 || pLhs->iTable!=iCur 2356 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2357 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2358 ){ 2359 break; 2360 } 2361 2362 testcase( pLhs->iColumn==XN_ROWID ); 2363 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2364 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2365 if( aff!=idxaff ) break; 2366 2367 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2368 if( pColl==0 ) break; 2369 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2370 } 2371 return i; 2372 } 2373 2374 /* 2375 ** Adjust the cost C by the costMult facter T. This only occurs if 2376 ** compiled with -DSQLITE_ENABLE_COSTMULT 2377 */ 2378 #ifdef SQLITE_ENABLE_COSTMULT 2379 # define ApplyCostMultiplier(C,T) C += T 2380 #else 2381 # define ApplyCostMultiplier(C,T) 2382 #endif 2383 2384 /* 2385 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2386 ** index pIndex. Try to match one more. 2387 ** 2388 ** When this function is called, pBuilder->pNew->nOut contains the 2389 ** number of rows expected to be visited by filtering using the nEq 2390 ** terms only. If it is modified, this value is restored before this 2391 ** function returns. 2392 ** 2393 ** If pProbe->idxType==SQLITE_IDXTYPE_IPK, that means pIndex is 2394 ** a fake index used for the INTEGER PRIMARY KEY. 2395 */ 2396 static int whereLoopAddBtreeIndex( 2397 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2398 struct SrcList_item *pSrc, /* FROM clause term being analyzed */ 2399 Index *pProbe, /* An index on pSrc */ 2400 LogEst nInMul /* log(Number of iterations due to IN) */ 2401 ){ 2402 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2403 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2404 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2405 WhereLoop *pNew; /* Template WhereLoop under construction */ 2406 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2407 int opMask; /* Valid operators for constraints */ 2408 WhereScan scan; /* Iterator for WHERE terms */ 2409 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2410 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2411 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2412 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2413 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2414 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2415 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2416 LogEst saved_nOut; /* Original value of pNew->nOut */ 2417 int rc = SQLITE_OK; /* Return code */ 2418 LogEst rSize; /* Number of rows in the table */ 2419 LogEst rLogSize; /* Logarithm of table size */ 2420 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2421 2422 pNew = pBuilder->pNew; 2423 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2424 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d\n", 2425 pProbe->pTable->zName,pProbe->zName, pNew->u.btree.nEq)); 2426 2427 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2428 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2429 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2430 opMask = WO_LT|WO_LE; 2431 }else{ 2432 assert( pNew->u.btree.nBtm==0 ); 2433 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2434 } 2435 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2436 2437 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2438 2439 saved_nEq = pNew->u.btree.nEq; 2440 saved_nBtm = pNew->u.btree.nBtm; 2441 saved_nTop = pNew->u.btree.nTop; 2442 saved_nSkip = pNew->nSkip; 2443 saved_nLTerm = pNew->nLTerm; 2444 saved_wsFlags = pNew->wsFlags; 2445 saved_prereq = pNew->prereq; 2446 saved_nOut = pNew->nOut; 2447 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2448 opMask, pProbe); 2449 pNew->rSetup = 0; 2450 rSize = pProbe->aiRowLogEst[0]; 2451 rLogSize = estLog(rSize); 2452 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2453 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2454 LogEst rCostIdx; 2455 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2456 int nIn = 0; 2457 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2458 int nRecValid = pBuilder->nRecValid; 2459 #endif 2460 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2461 && indexColumnNotNull(pProbe, saved_nEq) 2462 ){ 2463 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2464 } 2465 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2466 2467 /* Do not allow the upper bound of a LIKE optimization range constraint 2468 ** to mix with a lower range bound from some other source */ 2469 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2470 2471 /* Do not allow constraints from the WHERE clause to be used by the 2472 ** right table of a LEFT JOIN. Only constraints in the ON clause are 2473 ** allowed */ 2474 if( (pSrc->fg.jointype & JT_LEFT)!=0 2475 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2476 ){ 2477 continue; 2478 } 2479 2480 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2481 pBuilder->bldFlags |= SQLITE_BLDF_UNIQUE; 2482 }else{ 2483 pBuilder->bldFlags |= SQLITE_BLDF_INDEXED; 2484 } 2485 pNew->wsFlags = saved_wsFlags; 2486 pNew->u.btree.nEq = saved_nEq; 2487 pNew->u.btree.nBtm = saved_nBtm; 2488 pNew->u.btree.nTop = saved_nTop; 2489 pNew->nLTerm = saved_nLTerm; 2490 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2491 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2492 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2493 2494 assert( nInMul==0 2495 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2496 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2497 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2498 ); 2499 2500 if( eOp & WO_IN ){ 2501 Expr *pExpr = pTerm->pExpr; 2502 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 2503 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2504 int i; 2505 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2506 2507 /* The expression may actually be of the form (x, y) IN (SELECT...). 2508 ** In this case there is a separate term for each of (x) and (y). 2509 ** However, the nIn multiplier should only be applied once, not once 2510 ** for each such term. The following loop checks that pTerm is the 2511 ** first such term in use, and sets nIn back to 0 if it is not. */ 2512 for(i=0; i<pNew->nLTerm-1; i++){ 2513 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2514 } 2515 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2516 /* "x IN (value, value, ...)" */ 2517 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2518 assert( nIn>0 ); /* RHS always has 2 or more terms... The parser 2519 ** changes "x IN (?)" into "x=?". */ 2520 } 2521 if( pProbe->hasStat1 ){ 2522 LogEst M, logK, safetyMargin; 2523 /* Let: 2524 ** N = the total number of rows in the table 2525 ** K = the number of entries on the RHS of the IN operator 2526 ** M = the number of rows in the table that match terms to the 2527 ** to the left in the same index. If the IN operator is on 2528 ** the left-most index column, M==N. 2529 ** 2530 ** Given the definitions above, it is better to omit the IN operator 2531 ** from the index lookup and instead do a scan of the M elements, 2532 ** testing each scanned row against the IN operator separately, if: 2533 ** 2534 ** M*log(K) < K*log(N) 2535 ** 2536 ** Our estimates for M, K, and N might be inaccurate, so we build in 2537 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator 2538 ** with the index, as using an index has better worst-case behavior. 2539 ** If we do not have real sqlite_stat1 data, always prefer to use 2540 ** the index. 2541 */ 2542 M = pProbe->aiRowLogEst[saved_nEq]; 2543 logK = estLog(nIn); 2544 safetyMargin = 10; /* TUNING: extra weight for indexed IN */ 2545 if( M + logK + safetyMargin < nIn + rLogSize ){ 2546 WHERETRACE(0x40, 2547 ("Scan preferred over IN operator on column %d of \"%s\" (%d<%d)\n", 2548 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2549 continue; 2550 }else{ 2551 WHERETRACE(0x40, 2552 ("IN operator preferred on column %d of \"%s\" (%d>=%d)\n", 2553 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2554 } 2555 } 2556 pNew->wsFlags |= WHERE_COLUMN_IN; 2557 }else if( eOp & (WO_EQ|WO_IS) ){ 2558 int iCol = pProbe->aiColumn[saved_nEq]; 2559 pNew->wsFlags |= WHERE_COLUMN_EQ; 2560 assert( saved_nEq==pNew->u.btree.nEq ); 2561 if( iCol==XN_ROWID 2562 || (iCol>=0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2563 ){ 2564 if( iCol==XN_ROWID || pProbe->uniqNotNull 2565 || (pProbe->nKeyCol==1 && pProbe->onError && eOp==WO_EQ) 2566 ){ 2567 pNew->wsFlags |= WHERE_ONEROW; 2568 }else{ 2569 pNew->wsFlags |= WHERE_UNQ_WANTED; 2570 } 2571 } 2572 }else if( eOp & WO_ISNULL ){ 2573 pNew->wsFlags |= WHERE_COLUMN_NULL; 2574 }else if( eOp & (WO_GT|WO_GE) ){ 2575 testcase( eOp & WO_GT ); 2576 testcase( eOp & WO_GE ); 2577 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2578 pNew->u.btree.nBtm = whereRangeVectorLen( 2579 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2580 ); 2581 pBtm = pTerm; 2582 pTop = 0; 2583 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2584 /* Range contraints that come from the LIKE optimization are 2585 ** always used in pairs. */ 2586 pTop = &pTerm[1]; 2587 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2588 assert( pTop->wtFlags & TERM_LIKEOPT ); 2589 assert( pTop->eOperator==WO_LT ); 2590 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2591 pNew->aLTerm[pNew->nLTerm++] = pTop; 2592 pNew->wsFlags |= WHERE_TOP_LIMIT; 2593 pNew->u.btree.nTop = 1; 2594 } 2595 }else{ 2596 assert( eOp & (WO_LT|WO_LE) ); 2597 testcase( eOp & WO_LT ); 2598 testcase( eOp & WO_LE ); 2599 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2600 pNew->u.btree.nTop = whereRangeVectorLen( 2601 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2602 ); 2603 pTop = pTerm; 2604 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2605 pNew->aLTerm[pNew->nLTerm-2] : 0; 2606 } 2607 2608 /* At this point pNew->nOut is set to the number of rows expected to 2609 ** be visited by the index scan before considering term pTerm, or the 2610 ** values of nIn and nInMul. In other words, assuming that all 2611 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2612 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2613 assert( pNew->nOut==saved_nOut ); 2614 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2615 /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4 2616 ** data, using some other estimate. */ 2617 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2618 }else{ 2619 int nEq = ++pNew->u.btree.nEq; 2620 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2621 2622 assert( pNew->nOut==saved_nOut ); 2623 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2624 assert( (eOp & WO_IN) || nIn==0 ); 2625 testcase( eOp & WO_IN ); 2626 pNew->nOut += pTerm->truthProb; 2627 pNew->nOut -= nIn; 2628 }else{ 2629 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2630 tRowcnt nOut = 0; 2631 if( nInMul==0 2632 && pProbe->nSample 2633 && pNew->u.btree.nEq<=pProbe->nSampleCol 2634 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) 2635 && OptimizationEnabled(db, SQLITE_Stat34) 2636 ){ 2637 Expr *pExpr = pTerm->pExpr; 2638 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2639 testcase( eOp & WO_EQ ); 2640 testcase( eOp & WO_IS ); 2641 testcase( eOp & WO_ISNULL ); 2642 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2643 }else{ 2644 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2645 } 2646 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2647 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2648 if( nOut ){ 2649 pNew->nOut = sqlite3LogEst(nOut); 2650 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2651 pNew->nOut -= nIn; 2652 } 2653 } 2654 if( nOut==0 ) 2655 #endif 2656 { 2657 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2658 if( eOp & WO_ISNULL ){ 2659 /* TUNING: If there is no likelihood() value, assume that a 2660 ** "col IS NULL" expression matches twice as many rows 2661 ** as (col=?). */ 2662 pNew->nOut += 10; 2663 } 2664 } 2665 } 2666 } 2667 2668 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2669 ** it to pNew->rRun, which is currently set to the cost of the index 2670 ** seek only. Then, if this is a non-covering index, add the cost of 2671 ** visiting the rows in the main table. */ 2672 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2673 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2674 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2675 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2676 } 2677 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2678 2679 nOutUnadjusted = pNew->nOut; 2680 pNew->rRun += nInMul + nIn; 2681 pNew->nOut += nInMul + nIn; 2682 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 2683 rc = whereLoopInsert(pBuilder, pNew); 2684 2685 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2686 pNew->nOut = saved_nOut; 2687 }else{ 2688 pNew->nOut = nOutUnadjusted; 2689 } 2690 2691 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 2692 && pNew->u.btree.nEq<pProbe->nColumn 2693 ){ 2694 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 2695 } 2696 pNew->nOut = saved_nOut; 2697 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2698 pBuilder->nRecValid = nRecValid; 2699 #endif 2700 } 2701 pNew->prereq = saved_prereq; 2702 pNew->u.btree.nEq = saved_nEq; 2703 pNew->u.btree.nBtm = saved_nBtm; 2704 pNew->u.btree.nTop = saved_nTop; 2705 pNew->nSkip = saved_nSkip; 2706 pNew->wsFlags = saved_wsFlags; 2707 pNew->nOut = saved_nOut; 2708 pNew->nLTerm = saved_nLTerm; 2709 2710 /* Consider using a skip-scan if there are no WHERE clause constraints 2711 ** available for the left-most terms of the index, and if the average 2712 ** number of repeats in the left-most terms is at least 18. 2713 ** 2714 ** The magic number 18 is selected on the basis that scanning 17 rows 2715 ** is almost always quicker than an index seek (even though if the index 2716 ** contains fewer than 2^17 rows we assume otherwise in other parts of 2717 ** the code). And, even if it is not, it should not be too much slower. 2718 ** On the other hand, the extra seeks could end up being significantly 2719 ** more expensive. */ 2720 assert( 42==sqlite3LogEst(18) ); 2721 if( saved_nEq==saved_nSkip 2722 && saved_nEq+1<pProbe->nKeyCol 2723 && pProbe->noSkipScan==0 2724 && OptimizationEnabled(db, SQLITE_SkipScan) 2725 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 2726 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 2727 ){ 2728 LogEst nIter; 2729 pNew->u.btree.nEq++; 2730 pNew->nSkip++; 2731 pNew->aLTerm[pNew->nLTerm++] = 0; 2732 pNew->wsFlags |= WHERE_SKIPSCAN; 2733 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 2734 pNew->nOut -= nIter; 2735 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 2736 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 2737 nIter += 5; 2738 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 2739 pNew->nOut = saved_nOut; 2740 pNew->u.btree.nEq = saved_nEq; 2741 pNew->nSkip = saved_nSkip; 2742 pNew->wsFlags = saved_wsFlags; 2743 } 2744 2745 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n", 2746 pProbe->pTable->zName, pProbe->zName, saved_nEq, rc)); 2747 return rc; 2748 } 2749 2750 /* 2751 ** Return True if it is possible that pIndex might be useful in 2752 ** implementing the ORDER BY clause in pBuilder. 2753 ** 2754 ** Return False if pBuilder does not contain an ORDER BY clause or 2755 ** if there is no way for pIndex to be useful in implementing that 2756 ** ORDER BY clause. 2757 */ 2758 static int indexMightHelpWithOrderBy( 2759 WhereLoopBuilder *pBuilder, 2760 Index *pIndex, 2761 int iCursor 2762 ){ 2763 ExprList *pOB; 2764 ExprList *aColExpr; 2765 int ii, jj; 2766 2767 if( pIndex->bUnordered ) return 0; 2768 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 2769 for(ii=0; ii<pOB->nExpr; ii++){ 2770 Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr); 2771 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 2772 if( pExpr->iColumn<0 ) return 1; 2773 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2774 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 2775 } 2776 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 2777 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2778 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 2779 if( sqlite3ExprCompareSkip(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 2780 return 1; 2781 } 2782 } 2783 } 2784 } 2785 return 0; 2786 } 2787 2788 /* Check to see if a partial index with pPartIndexWhere can be used 2789 ** in the current query. Return true if it can be and false if not. 2790 */ 2791 static int whereUsablePartialIndex(int iTab, WhereClause *pWC, Expr *pWhere){ 2792 int i; 2793 WhereTerm *pTerm; 2794 Parse *pParse = pWC->pWInfo->pParse; 2795 while( pWhere->op==TK_AND ){ 2796 if( !whereUsablePartialIndex(iTab,pWC,pWhere->pLeft) ) return 0; 2797 pWhere = pWhere->pRight; 2798 } 2799 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 2800 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2801 Expr *pExpr = pTerm->pExpr; 2802 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 2803 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 2804 ){ 2805 return 1; 2806 } 2807 } 2808 return 0; 2809 } 2810 2811 /* 2812 ** Add all WhereLoop objects for a single table of the join where the table 2813 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 2814 ** a b-tree table, not a virtual table. 2815 ** 2816 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 2817 ** are calculated as follows: 2818 ** 2819 ** For a full scan, assuming the table (or index) contains nRow rows: 2820 ** 2821 ** cost = nRow * 3.0 // full-table scan 2822 ** cost = nRow * K // scan of covering index 2823 ** cost = nRow * (K+3.0) // scan of non-covering index 2824 ** 2825 ** where K is a value between 1.1 and 3.0 set based on the relative 2826 ** estimated average size of the index and table records. 2827 ** 2828 ** For an index scan, where nVisit is the number of index rows visited 2829 ** by the scan, and nSeek is the number of seek operations required on 2830 ** the index b-tree: 2831 ** 2832 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 2833 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 2834 ** 2835 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 2836 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 2837 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 2838 ** 2839 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 2840 ** of uncertainty. For this reason, scoring is designed to pick plans that 2841 ** "do the least harm" if the estimates are inaccurate. For example, a 2842 ** log(nRow) factor is omitted from a non-covering index scan in order to 2843 ** bias the scoring in favor of using an index, since the worst-case 2844 ** performance of using an index is far better than the worst-case performance 2845 ** of a full table scan. 2846 */ 2847 static int whereLoopAddBtree( 2848 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2849 Bitmask mPrereq /* Extra prerequesites for using this table */ 2850 ){ 2851 WhereInfo *pWInfo; /* WHERE analysis context */ 2852 Index *pProbe; /* An index we are evaluating */ 2853 Index sPk; /* A fake index object for the primary key */ 2854 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 2855 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 2856 SrcList *pTabList; /* The FROM clause */ 2857 struct SrcList_item *pSrc; /* The FROM clause btree term to add */ 2858 WhereLoop *pNew; /* Template WhereLoop object */ 2859 int rc = SQLITE_OK; /* Return code */ 2860 int iSortIdx = 1; /* Index number */ 2861 int b; /* A boolean value */ 2862 LogEst rSize; /* number of rows in the table */ 2863 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 2864 WhereClause *pWC; /* The parsed WHERE clause */ 2865 Table *pTab; /* Table being queried */ 2866 2867 pNew = pBuilder->pNew; 2868 pWInfo = pBuilder->pWInfo; 2869 pTabList = pWInfo->pTabList; 2870 pSrc = pTabList->a + pNew->iTab; 2871 pTab = pSrc->pTab; 2872 pWC = pBuilder->pWC; 2873 assert( !IsVirtual(pSrc->pTab) ); 2874 2875 if( pSrc->pIBIndex ){ 2876 /* An INDEXED BY clause specifies a particular index to use */ 2877 pProbe = pSrc->pIBIndex; 2878 }else if( !HasRowid(pTab) ){ 2879 pProbe = pTab->pIndex; 2880 }else{ 2881 /* There is no INDEXED BY clause. Create a fake Index object in local 2882 ** variable sPk to represent the rowid primary key index. Make this 2883 ** fake index the first in a chain of Index objects with all of the real 2884 ** indices to follow */ 2885 Index *pFirst; /* First of real indices on the table */ 2886 memset(&sPk, 0, sizeof(Index)); 2887 sPk.nKeyCol = 1; 2888 sPk.nColumn = 1; 2889 sPk.aiColumn = &aiColumnPk; 2890 sPk.aiRowLogEst = aiRowEstPk; 2891 sPk.onError = OE_Replace; 2892 sPk.pTable = pTab; 2893 sPk.szIdxRow = pTab->szTabRow; 2894 sPk.idxType = SQLITE_IDXTYPE_IPK; 2895 aiRowEstPk[0] = pTab->nRowLogEst; 2896 aiRowEstPk[1] = 0; 2897 pFirst = pSrc->pTab->pIndex; 2898 if( pSrc->fg.notIndexed==0 ){ 2899 /* The real indices of the table are only considered if the 2900 ** NOT INDEXED qualifier is omitted from the FROM clause */ 2901 sPk.pNext = pFirst; 2902 } 2903 pProbe = &sPk; 2904 } 2905 rSize = pTab->nRowLogEst; 2906 rLogSize = estLog(rSize); 2907 2908 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 2909 /* Automatic indexes */ 2910 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 2911 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 2912 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 2913 && pSrc->pIBIndex==0 /* Has no INDEXED BY clause */ 2914 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 2915 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 2916 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 2917 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 2918 ){ 2919 /* Generate auto-index WhereLoops */ 2920 WhereTerm *pTerm; 2921 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 2922 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 2923 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2924 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 2925 pNew->u.btree.nEq = 1; 2926 pNew->nSkip = 0; 2927 pNew->u.btree.pIndex = 0; 2928 pNew->nLTerm = 1; 2929 pNew->aLTerm[0] = pTerm; 2930 /* TUNING: One-time cost for computing the automatic index is 2931 ** estimated to be X*N*log2(N) where N is the number of rows in 2932 ** the table being indexed and where X is 7 (LogEst=28) for normal 2933 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value 2934 ** of X is smaller for views and subqueries so that the query planner 2935 ** will be more aggressive about generating automatic indexes for 2936 ** those objects, since there is no opportunity to add schema 2937 ** indexes on subqueries and views. */ 2938 pNew->rSetup = rLogSize + rSize; 2939 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ 2940 pNew->rSetup += 28; 2941 }else{ 2942 pNew->rSetup -= 10; 2943 } 2944 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 2945 if( pNew->rSetup<0 ) pNew->rSetup = 0; 2946 /* TUNING: Each index lookup yields 20 rows in the table. This 2947 ** is more than the usual guess of 10 rows, since we have no way 2948 ** of knowing how selective the index will ultimately be. It would 2949 ** not be unreasonable to make this value much larger. */ 2950 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 2951 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 2952 pNew->wsFlags = WHERE_AUTO_INDEX; 2953 pNew->prereq = mPrereq | pTerm->prereqRight; 2954 rc = whereLoopInsert(pBuilder, pNew); 2955 } 2956 } 2957 } 2958 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 2959 2960 /* Loop over all indices. If there was an INDEXED BY clause, then only 2961 ** consider index pProbe. */ 2962 for(; rc==SQLITE_OK && pProbe; 2963 pProbe=(pSrc->pIBIndex ? 0 : pProbe->pNext), iSortIdx++ 2964 ){ 2965 if( pProbe->pPartIdxWhere!=0 2966 && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){ 2967 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 2968 continue; /* Partial index inappropriate for this query */ 2969 } 2970 if( pProbe->bNoQuery ) continue; 2971 rSize = pProbe->aiRowLogEst[0]; 2972 pNew->u.btree.nEq = 0; 2973 pNew->u.btree.nBtm = 0; 2974 pNew->u.btree.nTop = 0; 2975 pNew->nSkip = 0; 2976 pNew->nLTerm = 0; 2977 pNew->iSortIdx = 0; 2978 pNew->rSetup = 0; 2979 pNew->prereq = mPrereq; 2980 pNew->nOut = rSize; 2981 pNew->u.btree.pIndex = pProbe; 2982 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 2983 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 2984 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 2985 if( pProbe->idxType==SQLITE_IDXTYPE_IPK ){ 2986 /* Integer primary key index */ 2987 pNew->wsFlags = WHERE_IPK; 2988 2989 /* Full table scan */ 2990 pNew->iSortIdx = b ? iSortIdx : 0; 2991 /* TUNING: Cost of full table scan is (N*3.0). */ 2992 pNew->rRun = rSize + 16; 2993 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2994 whereLoopOutputAdjust(pWC, pNew, rSize); 2995 rc = whereLoopInsert(pBuilder, pNew); 2996 pNew->nOut = rSize; 2997 if( rc ) break; 2998 }else{ 2999 Bitmask m; 3000 if( pProbe->isCovering ){ 3001 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 3002 m = 0; 3003 }else{ 3004 m = pSrc->colUsed & pProbe->colNotIdxed; 3005 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 3006 } 3007 3008 /* Full scan via index */ 3009 if( b 3010 || !HasRowid(pTab) 3011 || pProbe->pPartIdxWhere!=0 3012 || ( m==0 3013 && pProbe->bUnordered==0 3014 && (pProbe->szIdxRow<pTab->szTabRow) 3015 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 3016 && sqlite3GlobalConfig.bUseCis 3017 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 3018 ) 3019 ){ 3020 pNew->iSortIdx = b ? iSortIdx : 0; 3021 3022 /* The cost of visiting the index rows is N*K, where K is 3023 ** between 1.1 and 3.0, depending on the relative sizes of the 3024 ** index and table rows. */ 3025 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 3026 if( m!=0 ){ 3027 /* If this is a non-covering index scan, add in the cost of 3028 ** doing table lookups. The cost will be 3x the number of 3029 ** lookups. Take into account WHERE clause terms that can be 3030 ** satisfied using just the index, and that do not require a 3031 ** table lookup. */ 3032 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 3033 int ii; 3034 int iCur = pSrc->iCursor; 3035 WhereClause *pWC2 = &pWInfo->sWC; 3036 for(ii=0; ii<pWC2->nTerm; ii++){ 3037 WhereTerm *pTerm = &pWC2->a[ii]; 3038 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 3039 break; 3040 } 3041 /* pTerm can be evaluated using just the index. So reduce 3042 ** the expected number of table lookups accordingly */ 3043 if( pTerm->truthProb<=0 ){ 3044 nLookup += pTerm->truthProb; 3045 }else{ 3046 nLookup--; 3047 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 3048 } 3049 } 3050 3051 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 3052 } 3053 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3054 whereLoopOutputAdjust(pWC, pNew, rSize); 3055 rc = whereLoopInsert(pBuilder, pNew); 3056 pNew->nOut = rSize; 3057 if( rc ) break; 3058 } 3059 } 3060 3061 pBuilder->bldFlags = 0; 3062 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 3063 if( pBuilder->bldFlags==SQLITE_BLDF_INDEXED ){ 3064 /* If a non-unique index is used, or if a prefix of the key for 3065 ** unique index is used (making the index functionally non-unique) 3066 ** then the sqlite_stat1 data becomes important for scoring the 3067 ** plan */ 3068 pTab->tabFlags |= TF_StatsUsed; 3069 } 3070 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 3071 sqlite3Stat4ProbeFree(pBuilder->pRec); 3072 pBuilder->nRecValid = 0; 3073 pBuilder->pRec = 0; 3074 #endif 3075 } 3076 return rc; 3077 } 3078 3079 #ifndef SQLITE_OMIT_VIRTUALTABLE 3080 3081 /* 3082 ** Argument pIdxInfo is already populated with all constraints that may 3083 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 3084 ** function marks a subset of those constraints usable, invokes the 3085 ** xBestIndex method and adds the returned plan to pBuilder. 3086 ** 3087 ** A constraint is marked usable if: 3088 ** 3089 ** * Argument mUsable indicates that its prerequisites are available, and 3090 ** 3091 ** * It is not one of the operators specified in the mExclude mask passed 3092 ** as the fourth argument (which in practice is either WO_IN or 0). 3093 ** 3094 ** Argument mPrereq is a mask of tables that must be scanned before the 3095 ** virtual table in question. These are added to the plans prerequisites 3096 ** before it is added to pBuilder. 3097 ** 3098 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3099 ** uses one or more WO_IN terms, or false otherwise. 3100 */ 3101 static int whereLoopAddVirtualOne( 3102 WhereLoopBuilder *pBuilder, 3103 Bitmask mPrereq, /* Mask of tables that must be used. */ 3104 Bitmask mUsable, /* Mask of usable tables */ 3105 u16 mExclude, /* Exclude terms using these operators */ 3106 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3107 u16 mNoOmit, /* Do not omit these constraints */ 3108 int *pbIn /* OUT: True if plan uses an IN(...) op */ 3109 ){ 3110 WhereClause *pWC = pBuilder->pWC; 3111 struct sqlite3_index_constraint *pIdxCons; 3112 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3113 int i; 3114 int mxTerm; 3115 int rc = SQLITE_OK; 3116 WhereLoop *pNew = pBuilder->pNew; 3117 Parse *pParse = pBuilder->pWInfo->pParse; 3118 struct SrcList_item *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3119 int nConstraint = pIdxInfo->nConstraint; 3120 3121 assert( (mUsable & mPrereq)==mPrereq ); 3122 *pbIn = 0; 3123 pNew->prereq = mPrereq; 3124 3125 /* Set the usable flag on the subset of constraints identified by 3126 ** arguments mUsable and mExclude. */ 3127 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3128 for(i=0; i<nConstraint; i++, pIdxCons++){ 3129 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3130 pIdxCons->usable = 0; 3131 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3132 && (pTerm->eOperator & mExclude)==0 3133 ){ 3134 pIdxCons->usable = 1; 3135 } 3136 } 3137 3138 /* Initialize the output fields of the sqlite3_index_info structure */ 3139 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3140 assert( pIdxInfo->needToFreeIdxStr==0 ); 3141 pIdxInfo->idxStr = 0; 3142 pIdxInfo->idxNum = 0; 3143 pIdxInfo->orderByConsumed = 0; 3144 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3145 pIdxInfo->estimatedRows = 25; 3146 pIdxInfo->idxFlags = 0; 3147 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3148 3149 /* Invoke the virtual table xBestIndex() method */ 3150 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3151 if( rc ){ 3152 if( rc==SQLITE_CONSTRAINT ){ 3153 /* If the xBestIndex method returns SQLITE_CONSTRAINT, that means 3154 ** that the particular combination of parameters provided is unusable. 3155 ** Make no entries in the loop table. 3156 */ 3157 WHERETRACE(0xffff, (" ^^^^--- non-viable plan rejected!\n")); 3158 return SQLITE_OK; 3159 } 3160 return rc; 3161 } 3162 3163 mxTerm = -1; 3164 assert( pNew->nLSlot>=nConstraint ); 3165 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 3166 pNew->u.vtab.omitMask = 0; 3167 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3168 for(i=0; i<nConstraint; i++, pIdxCons++){ 3169 int iTerm; 3170 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3171 WhereTerm *pTerm; 3172 int j = pIdxCons->iTermOffset; 3173 if( iTerm>=nConstraint 3174 || j<0 3175 || j>=pWC->nTerm 3176 || pNew->aLTerm[iTerm]!=0 3177 || pIdxCons->usable==0 3178 ){ 3179 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3180 testcase( pIdxInfo->needToFreeIdxStr ); 3181 return SQLITE_ERROR; 3182 } 3183 testcase( iTerm==nConstraint-1 ); 3184 testcase( j==0 ); 3185 testcase( j==pWC->nTerm-1 ); 3186 pTerm = &pWC->a[j]; 3187 pNew->prereq |= pTerm->prereqRight; 3188 assert( iTerm<pNew->nLSlot ); 3189 pNew->aLTerm[iTerm] = pTerm; 3190 if( iTerm>mxTerm ) mxTerm = iTerm; 3191 testcase( iTerm==15 ); 3192 testcase( iTerm==16 ); 3193 if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<<iTerm; 3194 if( (pTerm->eOperator & WO_IN)!=0 ){ 3195 /* A virtual table that is constrained by an IN clause may not 3196 ** consume the ORDER BY clause because (1) the order of IN terms 3197 ** is not necessarily related to the order of output terms and 3198 ** (2) Multiple outputs from a single IN value will not merge 3199 ** together. */ 3200 pIdxInfo->orderByConsumed = 0; 3201 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3202 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3203 } 3204 } 3205 } 3206 pNew->u.vtab.omitMask &= ~mNoOmit; 3207 3208 pNew->nLTerm = mxTerm+1; 3209 for(i=0; i<=mxTerm; i++){ 3210 if( pNew->aLTerm[i]==0 ){ 3211 /* The non-zero argvIdx values must be contiguous. Raise an 3212 ** error if they are not */ 3213 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3214 testcase( pIdxInfo->needToFreeIdxStr ); 3215 return SQLITE_ERROR; 3216 } 3217 } 3218 assert( pNew->nLTerm<=pNew->nLSlot ); 3219 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3220 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3221 pIdxInfo->needToFreeIdxStr = 0; 3222 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3223 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3224 pIdxInfo->nOrderBy : 0); 3225 pNew->rSetup = 0; 3226 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3227 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3228 3229 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3230 ** that the scan will visit at most one row. Clear it otherwise. */ 3231 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3232 pNew->wsFlags |= WHERE_ONEROW; 3233 }else{ 3234 pNew->wsFlags &= ~WHERE_ONEROW; 3235 } 3236 rc = whereLoopInsert(pBuilder, pNew); 3237 if( pNew->u.vtab.needFree ){ 3238 sqlite3_free(pNew->u.vtab.idxStr); 3239 pNew->u.vtab.needFree = 0; 3240 } 3241 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3242 *pbIn, (sqlite3_uint64)mPrereq, 3243 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3244 3245 return rc; 3246 } 3247 3248 /* 3249 ** If this function is invoked from within an xBestIndex() callback, it 3250 ** returns a pointer to a buffer containing the name of the collation 3251 ** sequence associated with element iCons of the sqlite3_index_info.aConstraint 3252 ** array. Or, if iCons is out of range or there is no active xBestIndex 3253 ** call, return NULL. 3254 */ 3255 const char *sqlite3_vtab_collation(sqlite3_index_info *pIdxInfo, int iCons){ 3256 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3257 const char *zRet = 0; 3258 if( iCons>=0 && iCons<pIdxInfo->nConstraint ){ 3259 CollSeq *pC = 0; 3260 int iTerm = pIdxInfo->aConstraint[iCons].iTermOffset; 3261 Expr *pX = pHidden->pWC->a[iTerm].pExpr; 3262 if( pX->pLeft ){ 3263 pC = sqlite3BinaryCompareCollSeq(pHidden->pParse, pX->pLeft, pX->pRight); 3264 } 3265 zRet = (pC ? pC->zName : sqlite3StrBINARY); 3266 } 3267 return zRet; 3268 } 3269 3270 /* 3271 ** Add all WhereLoop objects for a table of the join identified by 3272 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3273 ** 3274 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3275 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3276 ** entries that occur before the virtual table in the FROM clause and are 3277 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3278 ** mUnusable mask contains all FROM clause entries that occur after the 3279 ** virtual table and are separated from it by at least one LEFT or 3280 ** CROSS JOIN. 3281 ** 3282 ** For example, if the query were: 3283 ** 3284 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3285 ** 3286 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3287 ** 3288 ** All the tables in mPrereq must be scanned before the current virtual 3289 ** table. So any terms for which all prerequisites are satisfied by 3290 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3291 ** Conversely, all tables in mUnusable must be scanned after the current 3292 ** virtual table, so any terms for which the prerequisites overlap with 3293 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3294 */ 3295 static int whereLoopAddVirtual( 3296 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3297 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3298 Bitmask mUnusable /* Tables that must be scanned after this one */ 3299 ){ 3300 int rc = SQLITE_OK; /* Return code */ 3301 WhereInfo *pWInfo; /* WHERE analysis context */ 3302 Parse *pParse; /* The parsing context */ 3303 WhereClause *pWC; /* The WHERE clause */ 3304 struct SrcList_item *pSrc; /* The FROM clause term to search */ 3305 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3306 int nConstraint; /* Number of constraints in p */ 3307 int bIn; /* True if plan uses IN(...) operator */ 3308 WhereLoop *pNew; 3309 Bitmask mBest; /* Tables used by best possible plan */ 3310 u16 mNoOmit; 3311 3312 assert( (mPrereq & mUnusable)==0 ); 3313 pWInfo = pBuilder->pWInfo; 3314 pParse = pWInfo->pParse; 3315 pWC = pBuilder->pWC; 3316 pNew = pBuilder->pNew; 3317 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3318 assert( IsVirtual(pSrc->pTab) ); 3319 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy, 3320 &mNoOmit); 3321 if( p==0 ) return SQLITE_NOMEM_BKPT; 3322 pNew->rSetup = 0; 3323 pNew->wsFlags = WHERE_VIRTUALTABLE; 3324 pNew->nLTerm = 0; 3325 pNew->u.vtab.needFree = 0; 3326 nConstraint = p->nConstraint; 3327 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3328 sqlite3DbFree(pParse->db, p); 3329 return SQLITE_NOMEM_BKPT; 3330 } 3331 3332 /* First call xBestIndex() with all constraints usable. */ 3333 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc->pTab->zName)); 3334 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3335 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn); 3336 3337 /* If the call to xBestIndex() with all terms enabled produced a plan 3338 ** that does not require any source tables (IOW: a plan with mBest==0) 3339 ** and does not use an IN(...) operator, then there is no point in making 3340 ** any further calls to xBestIndex() since they will all return the same 3341 ** result (if the xBestIndex() implementation is sane). */ 3342 if( rc==SQLITE_OK && ((mBest = (pNew->prereq & ~mPrereq))!=0 || bIn) ){ 3343 int seenZero = 0; /* True if a plan with no prereqs seen */ 3344 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3345 Bitmask mPrev = 0; 3346 Bitmask mBestNoIn = 0; 3347 3348 /* If the plan produced by the earlier call uses an IN(...) term, call 3349 ** xBestIndex again, this time with IN(...) terms disabled. */ 3350 if( bIn ){ 3351 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3352 rc = whereLoopAddVirtualOne( 3353 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn); 3354 assert( bIn==0 ); 3355 mBestNoIn = pNew->prereq & ~mPrereq; 3356 if( mBestNoIn==0 ){ 3357 seenZero = 1; 3358 seenZeroNoIN = 1; 3359 } 3360 } 3361 3362 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3363 ** in the set of terms that apply to the current virtual table. */ 3364 while( rc==SQLITE_OK ){ 3365 int i; 3366 Bitmask mNext = ALLBITS; 3367 assert( mNext>0 ); 3368 for(i=0; i<nConstraint; i++){ 3369 Bitmask mThis = ( 3370 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3371 ); 3372 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3373 } 3374 mPrev = mNext; 3375 if( mNext==ALLBITS ) break; 3376 if( mNext==mBest || mNext==mBestNoIn ) continue; 3377 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3378 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3379 rc = whereLoopAddVirtualOne( 3380 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn); 3381 if( pNew->prereq==mPrereq ){ 3382 seenZero = 1; 3383 if( bIn==0 ) seenZeroNoIN = 1; 3384 } 3385 } 3386 3387 /* If the calls to xBestIndex() in the above loop did not find a plan 3388 ** that requires no source tables at all (i.e. one guaranteed to be 3389 ** usable), make a call here with all source tables disabled */ 3390 if( rc==SQLITE_OK && seenZero==0 ){ 3391 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3392 rc = whereLoopAddVirtualOne( 3393 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn); 3394 if( bIn==0 ) seenZeroNoIN = 1; 3395 } 3396 3397 /* If the calls to xBestIndex() have so far failed to find a plan 3398 ** that requires no source tables at all and does not use an IN(...) 3399 ** operator, make a final call to obtain one here. */ 3400 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3401 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3402 rc = whereLoopAddVirtualOne( 3403 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn); 3404 } 3405 } 3406 3407 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3408 sqlite3DbFreeNN(pParse->db, p); 3409 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc->pTab->zName, rc)); 3410 return rc; 3411 } 3412 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3413 3414 /* 3415 ** Add WhereLoop entries to handle OR terms. This works for either 3416 ** btrees or virtual tables. 3417 */ 3418 static int whereLoopAddOr( 3419 WhereLoopBuilder *pBuilder, 3420 Bitmask mPrereq, 3421 Bitmask mUnusable 3422 ){ 3423 WhereInfo *pWInfo = pBuilder->pWInfo; 3424 WhereClause *pWC; 3425 WhereLoop *pNew; 3426 WhereTerm *pTerm, *pWCEnd; 3427 int rc = SQLITE_OK; 3428 int iCur; 3429 WhereClause tempWC; 3430 WhereLoopBuilder sSubBuild; 3431 WhereOrSet sSum, sCur; 3432 struct SrcList_item *pItem; 3433 3434 pWC = pBuilder->pWC; 3435 pWCEnd = pWC->a + pWC->nTerm; 3436 pNew = pBuilder->pNew; 3437 memset(&sSum, 0, sizeof(sSum)); 3438 pItem = pWInfo->pTabList->a + pNew->iTab; 3439 iCur = pItem->iCursor; 3440 3441 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3442 if( (pTerm->eOperator & WO_OR)!=0 3443 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3444 ){ 3445 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3446 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3447 WhereTerm *pOrTerm; 3448 int once = 1; 3449 int i, j; 3450 3451 sSubBuild = *pBuilder; 3452 sSubBuild.pOrderBy = 0; 3453 sSubBuild.pOrSet = &sCur; 3454 3455 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3456 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3457 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3458 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3459 }else if( pOrTerm->leftCursor==iCur ){ 3460 tempWC.pWInfo = pWC->pWInfo; 3461 tempWC.pOuter = pWC; 3462 tempWC.op = TK_AND; 3463 tempWC.nTerm = 1; 3464 tempWC.a = pOrTerm; 3465 sSubBuild.pWC = &tempWC; 3466 }else{ 3467 continue; 3468 } 3469 sCur.n = 0; 3470 #ifdef WHERETRACE_ENABLED 3471 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3472 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3473 if( sqlite3WhereTrace & 0x400 ){ 3474 sqlite3WhereClausePrint(sSubBuild.pWC); 3475 } 3476 #endif 3477 #ifndef SQLITE_OMIT_VIRTUALTABLE 3478 if( IsVirtual(pItem->pTab) ){ 3479 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 3480 }else 3481 #endif 3482 { 3483 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 3484 } 3485 if( rc==SQLITE_OK ){ 3486 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 3487 } 3488 assert( rc==SQLITE_OK || sCur.n==0 ); 3489 if( sCur.n==0 ){ 3490 sSum.n = 0; 3491 break; 3492 }else if( once ){ 3493 whereOrMove(&sSum, &sCur); 3494 once = 0; 3495 }else{ 3496 WhereOrSet sPrev; 3497 whereOrMove(&sPrev, &sSum); 3498 sSum.n = 0; 3499 for(i=0; i<sPrev.n; i++){ 3500 for(j=0; j<sCur.n; j++){ 3501 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3502 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3503 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3504 } 3505 } 3506 } 3507 } 3508 pNew->nLTerm = 1; 3509 pNew->aLTerm[0] = pTerm; 3510 pNew->wsFlags = WHERE_MULTI_OR; 3511 pNew->rSetup = 0; 3512 pNew->iSortIdx = 0; 3513 memset(&pNew->u, 0, sizeof(pNew->u)); 3514 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3515 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3516 ** of all sub-scans required by the OR-scan. However, due to rounding 3517 ** errors, it may be that the cost of the OR-scan is equal to its 3518 ** most expensive sub-scan. Add the smallest possible penalty 3519 ** (equivalent to multiplying the cost by 1.07) to ensure that 3520 ** this does not happen. Otherwise, for WHERE clauses such as the 3521 ** following where there is an index on "y": 3522 ** 3523 ** WHERE likelihood(x=?, 0.99) OR y=? 3524 ** 3525 ** the planner may elect to "OR" together a full-table scan and an 3526 ** index lookup. And other similarly odd results. */ 3527 pNew->rRun = sSum.a[i].rRun + 1; 3528 pNew->nOut = sSum.a[i].nOut; 3529 pNew->prereq = sSum.a[i].prereq; 3530 rc = whereLoopInsert(pBuilder, pNew); 3531 } 3532 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3533 } 3534 } 3535 return rc; 3536 } 3537 3538 /* 3539 ** Add all WhereLoop objects for all tables 3540 */ 3541 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3542 WhereInfo *pWInfo = pBuilder->pWInfo; 3543 Bitmask mPrereq = 0; 3544 Bitmask mPrior = 0; 3545 int iTab; 3546 SrcList *pTabList = pWInfo->pTabList; 3547 struct SrcList_item *pItem; 3548 struct SrcList_item *pEnd = &pTabList->a[pWInfo->nLevel]; 3549 sqlite3 *db = pWInfo->pParse->db; 3550 int rc = SQLITE_OK; 3551 WhereLoop *pNew; 3552 u8 priorJointype = 0; 3553 3554 /* Loop over the tables in the join, from left to right */ 3555 pNew = pBuilder->pNew; 3556 whereLoopInit(pNew); 3557 pBuilder->iPlanLimit = SQLITE_QUERY_PLANNER_LIMIT; 3558 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3559 Bitmask mUnusable = 0; 3560 pNew->iTab = iTab; 3561 pBuilder->iPlanLimit += SQLITE_QUERY_PLANNER_LIMIT_INCR; 3562 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3563 if( ((pItem->fg.jointype|priorJointype) & (JT_LEFT|JT_CROSS))!=0 ){ 3564 /* This condition is true when pItem is the FROM clause term on the 3565 ** right-hand-side of a LEFT or CROSS JOIN. */ 3566 mPrereq = mPrior; 3567 } 3568 priorJointype = pItem->fg.jointype; 3569 #ifndef SQLITE_OMIT_VIRTUALTABLE 3570 if( IsVirtual(pItem->pTab) ){ 3571 struct SrcList_item *p; 3572 for(p=&pItem[1]; p<pEnd; p++){ 3573 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3574 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3575 } 3576 } 3577 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 3578 }else 3579 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3580 { 3581 rc = whereLoopAddBtree(pBuilder, mPrereq); 3582 } 3583 if( rc==SQLITE_OK && pBuilder->pWC->hasOr ){ 3584 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 3585 } 3586 mPrior |= pNew->maskSelf; 3587 if( rc || db->mallocFailed ){ 3588 if( rc==SQLITE_DONE ){ 3589 /* We hit the query planner search limit set by iPlanLimit */ 3590 sqlite3_log(SQLITE_WARNING, "abbreviated query algorithm search"); 3591 rc = SQLITE_OK; 3592 }else{ 3593 break; 3594 } 3595 } 3596 } 3597 3598 whereLoopClear(db, pNew); 3599 return rc; 3600 } 3601 3602 /* 3603 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 3604 ** parameters) to see if it outputs rows in the requested ORDER BY 3605 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3606 ** 3607 ** N>0: N terms of the ORDER BY clause are satisfied 3608 ** N==0: No terms of the ORDER BY clause are satisfied 3609 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3610 ** 3611 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3612 ** strict. With GROUP BY and DISTINCT the only requirement is that 3613 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3614 ** and DISTINCT do not require rows to appear in any particular order as long 3615 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3616 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3617 ** pOrderBy terms must be matched in strict left-to-right order. 3618 */ 3619 static i8 wherePathSatisfiesOrderBy( 3620 WhereInfo *pWInfo, /* The WHERE clause */ 3621 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3622 WherePath *pPath, /* The WherePath to check */ 3623 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 3624 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3625 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3626 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3627 ){ 3628 u8 revSet; /* True if rev is known */ 3629 u8 rev; /* Composite sort order */ 3630 u8 revIdx; /* Index sort order */ 3631 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 3632 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 3633 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 3634 u16 eqOpMask; /* Allowed equality operators */ 3635 u16 nKeyCol; /* Number of key columns in pIndex */ 3636 u16 nColumn; /* Total number of ordered columns in the index */ 3637 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 3638 int iLoop; /* Index of WhereLoop in pPath being processed */ 3639 int i, j; /* Loop counters */ 3640 int iCur; /* Cursor number for current WhereLoop */ 3641 int iColumn; /* A column number within table iCur */ 3642 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 3643 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3644 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 3645 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 3646 Index *pIndex; /* The index associated with pLoop */ 3647 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 3648 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 3649 Bitmask obDone; /* Mask of all ORDER BY terms */ 3650 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 3651 Bitmask ready; /* Mask of inner loops */ 3652 3653 /* 3654 ** We say the WhereLoop is "one-row" if it generates no more than one 3655 ** row of output. A WhereLoop is one-row if all of the following are true: 3656 ** (a) All index columns match with WHERE_COLUMN_EQ. 3657 ** (b) The index is unique 3658 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 3659 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 3660 ** 3661 ** We say the WhereLoop is "order-distinct" if the set of columns from 3662 ** that WhereLoop that are in the ORDER BY clause are different for every 3663 ** row of the WhereLoop. Every one-row WhereLoop is automatically 3664 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 3665 ** is not order-distinct. To be order-distinct is not quite the same as being 3666 ** UNIQUE since a UNIQUE column or index can have multiple rows that 3667 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 3668 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 3669 ** 3670 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 3671 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 3672 ** automatically order-distinct. 3673 */ 3674 3675 assert( pOrderBy!=0 ); 3676 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 3677 3678 nOrderBy = pOrderBy->nExpr; 3679 testcase( nOrderBy==BMS-1 ); 3680 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 3681 isOrderDistinct = 1; 3682 obDone = MASKBIT(nOrderBy)-1; 3683 orderDistinctMask = 0; 3684 ready = 0; 3685 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 3686 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) eqOpMask |= WO_IN; 3687 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 3688 if( iLoop>0 ) ready |= pLoop->maskSelf; 3689 if( iLoop<nLoop ){ 3690 pLoop = pPath->aLoop[iLoop]; 3691 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 3692 }else{ 3693 pLoop = pLast; 3694 } 3695 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 3696 if( pLoop->u.vtab.isOrdered ) obSat = obDone; 3697 break; 3698 }else{ 3699 pLoop->u.btree.nIdxCol = 0; 3700 } 3701 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 3702 3703 /* Mark off any ORDER BY term X that is a column in the table of 3704 ** the current loop for which there is term in the WHERE 3705 ** clause of the form X IS NULL or X=? that reference only outer 3706 ** loops. 3707 */ 3708 for(i=0; i<nOrderBy; i++){ 3709 if( MASKBIT(i) & obSat ) continue; 3710 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3711 if( pOBExpr->op!=TK_COLUMN ) continue; 3712 if( pOBExpr->iTable!=iCur ) continue; 3713 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 3714 ~ready, eqOpMask, 0); 3715 if( pTerm==0 ) continue; 3716 if( pTerm->eOperator==WO_IN ){ 3717 /* IN terms are only valid for sorting in the ORDER BY LIMIT 3718 ** optimization, and then only if they are actually used 3719 ** by the query plan */ 3720 assert( wctrlFlags & WHERE_ORDERBY_LIMIT ); 3721 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 3722 if( j>=pLoop->nLTerm ) continue; 3723 } 3724 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 3725 if( sqlite3ExprCollSeqMatch(pWInfo->pParse, 3726 pOrderBy->a[i].pExpr, pTerm->pExpr)==0 ){ 3727 continue; 3728 } 3729 testcase( pTerm->pExpr->op==TK_IS ); 3730 } 3731 obSat |= MASKBIT(i); 3732 } 3733 3734 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 3735 if( pLoop->wsFlags & WHERE_IPK ){ 3736 pIndex = 0; 3737 nKeyCol = 0; 3738 nColumn = 1; 3739 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 3740 return 0; 3741 }else{ 3742 nKeyCol = pIndex->nKeyCol; 3743 nColumn = pIndex->nColumn; 3744 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 3745 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 3746 || !HasRowid(pIndex->pTable)); 3747 isOrderDistinct = IsUniqueIndex(pIndex); 3748 } 3749 3750 /* Loop through all columns of the index and deal with the ones 3751 ** that are not constrained by == or IN. 3752 */ 3753 rev = revSet = 0; 3754 distinctColumns = 0; 3755 for(j=0; j<nColumn; j++){ 3756 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 3757 3758 assert( j>=pLoop->u.btree.nEq 3759 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 3760 ); 3761 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 3762 u16 eOp = pLoop->aLTerm[j]->eOperator; 3763 3764 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 3765 ** doing WHERE_ORDERBY_LIMIT processing). 3766 ** 3767 ** If the current term is a column of an ((?,?) IN (SELECT...)) 3768 ** expression for which the SELECT returns more than one column, 3769 ** check that it is the only column used by this loop. Otherwise, 3770 ** if it is one of two or more, none of the columns can be 3771 ** considered to match an ORDER BY term. */ 3772 if( (eOp & eqOpMask)!=0 ){ 3773 if( eOp & WO_ISNULL ){ 3774 testcase( isOrderDistinct ); 3775 isOrderDistinct = 0; 3776 } 3777 continue; 3778 }else if( ALWAYS(eOp & WO_IN) ){ 3779 /* ALWAYS() justification: eOp is an equality operator due to the 3780 ** j<pLoop->u.btree.nEq constraint above. Any equality other 3781 ** than WO_IN is captured by the previous "if". So this one 3782 ** always has to be WO_IN. */ 3783 Expr *pX = pLoop->aLTerm[j]->pExpr; 3784 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 3785 if( pLoop->aLTerm[i]->pExpr==pX ){ 3786 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 3787 bOnce = 0; 3788 break; 3789 } 3790 } 3791 } 3792 } 3793 3794 /* Get the column number in the table (iColumn) and sort order 3795 ** (revIdx) for the j-th column of the index. 3796 */ 3797 if( pIndex ){ 3798 iColumn = pIndex->aiColumn[j]; 3799 revIdx = pIndex->aSortOrder[j]; 3800 if( iColumn==pIndex->pTable->iPKey ) iColumn = XN_ROWID; 3801 }else{ 3802 iColumn = XN_ROWID; 3803 revIdx = 0; 3804 } 3805 3806 /* An unconstrained column that might be NULL means that this 3807 ** WhereLoop is not well-ordered 3808 */ 3809 if( isOrderDistinct 3810 && iColumn>=0 3811 && j>=pLoop->u.btree.nEq 3812 && pIndex->pTable->aCol[iColumn].notNull==0 3813 ){ 3814 isOrderDistinct = 0; 3815 } 3816 3817 /* Find the ORDER BY term that corresponds to the j-th column 3818 ** of the index and mark that ORDER BY term off 3819 */ 3820 isMatch = 0; 3821 for(i=0; bOnce && i<nOrderBy; i++){ 3822 if( MASKBIT(i) & obSat ) continue; 3823 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3824 testcase( wctrlFlags & WHERE_GROUPBY ); 3825 testcase( wctrlFlags & WHERE_DISTINCTBY ); 3826 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 3827 if( iColumn>=XN_ROWID ){ 3828 if( pOBExpr->op!=TK_COLUMN ) continue; 3829 if( pOBExpr->iTable!=iCur ) continue; 3830 if( pOBExpr->iColumn!=iColumn ) continue; 3831 }else{ 3832 Expr *pIdxExpr = pIndex->aColExpr->a[j].pExpr; 3833 if( sqlite3ExprCompareSkip(pOBExpr, pIdxExpr, iCur) ){ 3834 continue; 3835 } 3836 } 3837 if( iColumn!=XN_ROWID ){ 3838 pColl = sqlite3ExprNNCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3839 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 3840 } 3841 pLoop->u.btree.nIdxCol = j+1; 3842 isMatch = 1; 3843 break; 3844 } 3845 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 3846 /* Make sure the sort order is compatible in an ORDER BY clause. 3847 ** Sort order is irrelevant for a GROUP BY clause. */ 3848 if( revSet ){ 3849 if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0; 3850 }else{ 3851 rev = revIdx ^ pOrderBy->a[i].sortOrder; 3852 if( rev ) *pRevMask |= MASKBIT(iLoop); 3853 revSet = 1; 3854 } 3855 } 3856 if( isMatch ){ 3857 if( iColumn==XN_ROWID ){ 3858 testcase( distinctColumns==0 ); 3859 distinctColumns = 1; 3860 } 3861 obSat |= MASKBIT(i); 3862 }else{ 3863 /* No match found */ 3864 if( j==0 || j<nKeyCol ){ 3865 testcase( isOrderDistinct!=0 ); 3866 isOrderDistinct = 0; 3867 } 3868 break; 3869 } 3870 } /* end Loop over all index columns */ 3871 if( distinctColumns ){ 3872 testcase( isOrderDistinct==0 ); 3873 isOrderDistinct = 1; 3874 } 3875 } /* end-if not one-row */ 3876 3877 /* Mark off any other ORDER BY terms that reference pLoop */ 3878 if( isOrderDistinct ){ 3879 orderDistinctMask |= pLoop->maskSelf; 3880 for(i=0; i<nOrderBy; i++){ 3881 Expr *p; 3882 Bitmask mTerm; 3883 if( MASKBIT(i) & obSat ) continue; 3884 p = pOrderBy->a[i].pExpr; 3885 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 3886 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 3887 if( (mTerm&~orderDistinctMask)==0 ){ 3888 obSat |= MASKBIT(i); 3889 } 3890 } 3891 } 3892 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 3893 if( obSat==obDone ) return (i8)nOrderBy; 3894 if( !isOrderDistinct ){ 3895 for(i=nOrderBy-1; i>0; i--){ 3896 Bitmask m = MASKBIT(i) - 1; 3897 if( (obSat&m)==m ) return i; 3898 } 3899 return 0; 3900 } 3901 return -1; 3902 } 3903 3904 3905 /* 3906 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 3907 ** the planner assumes that the specified pOrderBy list is actually a GROUP 3908 ** BY clause - and so any order that groups rows as required satisfies the 3909 ** request. 3910 ** 3911 ** Normally, in this case it is not possible for the caller to determine 3912 ** whether or not the rows are really being delivered in sorted order, or 3913 ** just in some other order that provides the required grouping. However, 3914 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 3915 ** this function may be called on the returned WhereInfo object. It returns 3916 ** true if the rows really will be sorted in the specified order, or false 3917 ** otherwise. 3918 ** 3919 ** For example, assuming: 3920 ** 3921 ** CREATE INDEX i1 ON t1(x, Y); 3922 ** 3923 ** then 3924 ** 3925 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 3926 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 3927 */ 3928 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 3929 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 3930 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 3931 return pWInfo->sorted; 3932 } 3933 3934 #ifdef WHERETRACE_ENABLED 3935 /* For debugging use only: */ 3936 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 3937 static char zName[65]; 3938 int i; 3939 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 3940 if( pLast ) zName[i++] = pLast->cId; 3941 zName[i] = 0; 3942 return zName; 3943 } 3944 #endif 3945 3946 /* 3947 ** Return the cost of sorting nRow rows, assuming that the keys have 3948 ** nOrderby columns and that the first nSorted columns are already in 3949 ** order. 3950 */ 3951 static LogEst whereSortingCost( 3952 WhereInfo *pWInfo, 3953 LogEst nRow, 3954 int nOrderBy, 3955 int nSorted 3956 ){ 3957 /* TUNING: Estimated cost of a full external sort, where N is 3958 ** the number of rows to sort is: 3959 ** 3960 ** cost = (3.0 * N * log(N)). 3961 ** 3962 ** Or, if the order-by clause has X terms but only the last Y 3963 ** terms are out of order, then block-sorting will reduce the 3964 ** sorting cost to: 3965 ** 3966 ** cost = (3.0 * N * log(N)) * (Y/X) 3967 ** 3968 ** The (Y/X) term is implemented using stack variable rScale 3969 ** below. */ 3970 LogEst rScale, rSortCost; 3971 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 3972 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 3973 rSortCost = nRow + rScale + 16; 3974 3975 /* Multiple by log(M) where M is the number of output rows. 3976 ** Use the LIMIT for M if it is smaller */ 3977 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 3978 nRow = pWInfo->iLimit; 3979 } 3980 rSortCost += estLog(nRow); 3981 return rSortCost; 3982 } 3983 3984 /* 3985 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 3986 ** attempts to find the lowest cost path that visits each WhereLoop 3987 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 3988 ** 3989 ** Assume that the total number of output rows that will need to be sorted 3990 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 3991 ** costs if nRowEst==0. 3992 ** 3993 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 3994 ** error occurs. 3995 */ 3996 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 3997 int mxChoice; /* Maximum number of simultaneous paths tracked */ 3998 int nLoop; /* Number of terms in the join */ 3999 Parse *pParse; /* Parsing context */ 4000 sqlite3 *db; /* The database connection */ 4001 int iLoop; /* Loop counter over the terms of the join */ 4002 int ii, jj; /* Loop counters */ 4003 int mxI = 0; /* Index of next entry to replace */ 4004 int nOrderBy; /* Number of ORDER BY clause terms */ 4005 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 4006 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 4007 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 4008 WherePath *aFrom; /* All nFrom paths at the previous level */ 4009 WherePath *aTo; /* The nTo best paths at the current level */ 4010 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 4011 WherePath *pTo; /* An element of aTo[] that we are working on */ 4012 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 4013 WhereLoop **pX; /* Used to divy up the pSpace memory */ 4014 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 4015 char *pSpace; /* Temporary memory used by this routine */ 4016 int nSpace; /* Bytes of space allocated at pSpace */ 4017 4018 pParse = pWInfo->pParse; 4019 db = pParse->db; 4020 nLoop = pWInfo->nLevel; 4021 /* TUNING: For simple queries, only the best path is tracked. 4022 ** For 2-way joins, the 5 best paths are followed. 4023 ** For joins of 3 or more tables, track the 10 best paths */ 4024 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 4025 assert( nLoop<=pWInfo->pTabList->nSrc ); 4026 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 4027 4028 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 4029 ** case the purpose of this call is to estimate the number of rows returned 4030 ** by the overall query. Once this estimate has been obtained, the caller 4031 ** will invoke this function a second time, passing the estimate as the 4032 ** nRowEst parameter. */ 4033 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 4034 nOrderBy = 0; 4035 }else{ 4036 nOrderBy = pWInfo->pOrderBy->nExpr; 4037 } 4038 4039 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 4040 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 4041 nSpace += sizeof(LogEst) * nOrderBy; 4042 pSpace = sqlite3DbMallocRawNN(db, nSpace); 4043 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 4044 aTo = (WherePath*)pSpace; 4045 aFrom = aTo+mxChoice; 4046 memset(aFrom, 0, sizeof(aFrom[0])); 4047 pX = (WhereLoop**)(aFrom+mxChoice); 4048 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 4049 pFrom->aLoop = pX; 4050 } 4051 if( nOrderBy ){ 4052 /* If there is an ORDER BY clause and it is not being ignored, set up 4053 ** space for the aSortCost[] array. Each element of the aSortCost array 4054 ** is either zero - meaning it has not yet been initialized - or the 4055 ** cost of sorting nRowEst rows of data where the first X terms of 4056 ** the ORDER BY clause are already in order, where X is the array 4057 ** index. */ 4058 aSortCost = (LogEst*)pX; 4059 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 4060 } 4061 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 4062 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 4063 4064 /* Seed the search with a single WherePath containing zero WhereLoops. 4065 ** 4066 ** TUNING: Do not let the number of iterations go above 28. If the cost 4067 ** of computing an automatic index is not paid back within the first 28 4068 ** rows, then do not use the automatic index. */ 4069 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 4070 nFrom = 1; 4071 assert( aFrom[0].isOrdered==0 ); 4072 if( nOrderBy ){ 4073 /* If nLoop is zero, then there are no FROM terms in the query. Since 4074 ** in this case the query may return a maximum of one row, the results 4075 ** are already in the requested order. Set isOrdered to nOrderBy to 4076 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 4077 ** -1, indicating that the result set may or may not be ordered, 4078 ** depending on the loops added to the current plan. */ 4079 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 4080 } 4081 4082 /* Compute successively longer WherePaths using the previous generation 4083 ** of WherePaths as the basis for the next. Keep track of the mxChoice 4084 ** best paths at each generation */ 4085 for(iLoop=0; iLoop<nLoop; iLoop++){ 4086 nTo = 0; 4087 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 4088 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 4089 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 4090 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 4091 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 4092 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 4093 Bitmask maskNew; /* Mask of src visited by (..) */ 4094 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 4095 4096 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 4097 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 4098 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<3 ){ 4099 /* Do not use an automatic index if the this loop is expected 4100 ** to run less than 1.25 times. It is tempting to also exclude 4101 ** automatic index usage on an outer loop, but sometimes an automatic 4102 ** index is useful in the outer loop of a correlated subquery. */ 4103 assert( 10==sqlite3LogEst(2) ); 4104 continue; 4105 } 4106 4107 /* At this point, pWLoop is a candidate to be the next loop. 4108 ** Compute its cost */ 4109 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 4110 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 4111 nOut = pFrom->nRow + pWLoop->nOut; 4112 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 4113 if( isOrdered<0 ){ 4114 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 4115 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 4116 iLoop, pWLoop, &revMask); 4117 }else{ 4118 revMask = pFrom->revLoop; 4119 } 4120 if( isOrdered>=0 && isOrdered<nOrderBy ){ 4121 if( aSortCost[isOrdered]==0 ){ 4122 aSortCost[isOrdered] = whereSortingCost( 4123 pWInfo, nRowEst, nOrderBy, isOrdered 4124 ); 4125 } 4126 /* TUNING: Add a small extra penalty (5) to sorting as an 4127 ** extra encouragment to the query planner to select a plan 4128 ** where the rows emerge in the correct order without any sorting 4129 ** required. */ 4130 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]) + 5; 4131 4132 WHERETRACE(0x002, 4133 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 4134 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 4135 rUnsorted, rCost)); 4136 }else{ 4137 rCost = rUnsorted; 4138 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 4139 } 4140 4141 /* Check to see if pWLoop should be added to the set of 4142 ** mxChoice best-so-far paths. 4143 ** 4144 ** First look for an existing path among best-so-far paths 4145 ** that covers the same set of loops and has the same isOrdered 4146 ** setting as the current path candidate. 4147 ** 4148 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 4149 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 4150 ** of legal values for isOrdered, -1..64. 4151 */ 4152 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4153 if( pTo->maskLoop==maskNew 4154 && ((pTo->isOrdered^isOrdered)&0x80)==0 4155 ){ 4156 testcase( jj==nTo-1 ); 4157 break; 4158 } 4159 } 4160 if( jj>=nTo ){ 4161 /* None of the existing best-so-far paths match the candidate. */ 4162 if( nTo>=mxChoice 4163 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4164 ){ 4165 /* The current candidate is no better than any of the mxChoice 4166 ** paths currently in the best-so-far buffer. So discard 4167 ** this candidate as not viable. */ 4168 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4169 if( sqlite3WhereTrace&0x4 ){ 4170 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4171 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4172 isOrdered>=0 ? isOrdered+'0' : '?'); 4173 } 4174 #endif 4175 continue; 4176 } 4177 /* If we reach this points it means that the new candidate path 4178 ** needs to be added to the set of best-so-far paths. */ 4179 if( nTo<mxChoice ){ 4180 /* Increase the size of the aTo set by one */ 4181 jj = nTo++; 4182 }else{ 4183 /* New path replaces the prior worst to keep count below mxChoice */ 4184 jj = mxI; 4185 } 4186 pTo = &aTo[jj]; 4187 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4188 if( sqlite3WhereTrace&0x4 ){ 4189 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4190 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4191 isOrdered>=0 ? isOrdered+'0' : '?'); 4192 } 4193 #endif 4194 }else{ 4195 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4196 ** same set of loops and has the same isOrdered setting as the 4197 ** candidate path. Check to see if the candidate should replace 4198 ** pTo or if the candidate should be skipped. 4199 ** 4200 ** The conditional is an expanded vector comparison equivalent to: 4201 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4202 */ 4203 if( pTo->rCost<rCost 4204 || (pTo->rCost==rCost 4205 && (pTo->nRow<nOut 4206 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4207 ) 4208 ) 4209 ){ 4210 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4211 if( sqlite3WhereTrace&0x4 ){ 4212 sqlite3DebugPrintf( 4213 "Skip %s cost=%-3d,%3d,%3d order=%c", 4214 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4215 isOrdered>=0 ? isOrdered+'0' : '?'); 4216 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4217 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4218 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4219 } 4220 #endif 4221 /* Discard the candidate path from further consideration */ 4222 testcase( pTo->rCost==rCost ); 4223 continue; 4224 } 4225 testcase( pTo->rCost==rCost+1 ); 4226 /* Control reaches here if the candidate path is better than the 4227 ** pTo path. Replace pTo with the candidate. */ 4228 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4229 if( sqlite3WhereTrace&0x4 ){ 4230 sqlite3DebugPrintf( 4231 "Update %s cost=%-3d,%3d,%3d order=%c", 4232 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4233 isOrdered>=0 ? isOrdered+'0' : '?'); 4234 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4235 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4236 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4237 } 4238 #endif 4239 } 4240 /* pWLoop is a winner. Add it to the set of best so far */ 4241 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4242 pTo->revLoop = revMask; 4243 pTo->nRow = nOut; 4244 pTo->rCost = rCost; 4245 pTo->rUnsorted = rUnsorted; 4246 pTo->isOrdered = isOrdered; 4247 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4248 pTo->aLoop[iLoop] = pWLoop; 4249 if( nTo>=mxChoice ){ 4250 mxI = 0; 4251 mxCost = aTo[0].rCost; 4252 mxUnsorted = aTo[0].nRow; 4253 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4254 if( pTo->rCost>mxCost 4255 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4256 ){ 4257 mxCost = pTo->rCost; 4258 mxUnsorted = pTo->rUnsorted; 4259 mxI = jj; 4260 } 4261 } 4262 } 4263 } 4264 } 4265 4266 #ifdef WHERETRACE_ENABLED /* >=2 */ 4267 if( sqlite3WhereTrace & 0x02 ){ 4268 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4269 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4270 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4271 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4272 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4273 if( pTo->isOrdered>0 ){ 4274 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4275 }else{ 4276 sqlite3DebugPrintf("\n"); 4277 } 4278 } 4279 } 4280 #endif 4281 4282 /* Swap the roles of aFrom and aTo for the next generation */ 4283 pFrom = aTo; 4284 aTo = aFrom; 4285 aFrom = pFrom; 4286 nFrom = nTo; 4287 } 4288 4289 if( nFrom==0 ){ 4290 sqlite3ErrorMsg(pParse, "no query solution"); 4291 sqlite3DbFreeNN(db, pSpace); 4292 return SQLITE_ERROR; 4293 } 4294 4295 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4296 pFrom = aFrom; 4297 for(ii=1; ii<nFrom; ii++){ 4298 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4299 } 4300 assert( pWInfo->nLevel==nLoop ); 4301 /* Load the lowest cost path into pWInfo */ 4302 for(iLoop=0; iLoop<nLoop; iLoop++){ 4303 WhereLevel *pLevel = pWInfo->a + iLoop; 4304 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4305 pLevel->iFrom = pWLoop->iTab; 4306 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4307 } 4308 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4309 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4310 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4311 && nRowEst 4312 ){ 4313 Bitmask notUsed; 4314 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4315 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4316 if( rc==pWInfo->pResultSet->nExpr ){ 4317 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4318 } 4319 } 4320 pWInfo->bOrderedInnerLoop = 0; 4321 if( pWInfo->pOrderBy ){ 4322 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4323 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4324 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4325 } 4326 }else{ 4327 pWInfo->nOBSat = pFrom->isOrdered; 4328 pWInfo->revMask = pFrom->revLoop; 4329 if( pWInfo->nOBSat<=0 ){ 4330 pWInfo->nOBSat = 0; 4331 if( nLoop>0 ){ 4332 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4333 if( (wsFlags & WHERE_ONEROW)==0 4334 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4335 ){ 4336 Bitmask m = 0; 4337 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4338 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4339 testcase( wsFlags & WHERE_IPK ); 4340 testcase( wsFlags & WHERE_COLUMN_IN ); 4341 if( rc==pWInfo->pOrderBy->nExpr ){ 4342 pWInfo->bOrderedInnerLoop = 1; 4343 pWInfo->revMask = m; 4344 } 4345 } 4346 } 4347 } 4348 } 4349 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4350 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4351 ){ 4352 Bitmask revMask = 0; 4353 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4354 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4355 ); 4356 assert( pWInfo->sorted==0 ); 4357 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4358 pWInfo->sorted = 1; 4359 pWInfo->revMask = revMask; 4360 } 4361 } 4362 } 4363 4364 4365 pWInfo->nRowOut = pFrom->nRow; 4366 4367 /* Free temporary memory and return success */ 4368 sqlite3DbFreeNN(db, pSpace); 4369 return SQLITE_OK; 4370 } 4371 4372 /* 4373 ** Most queries use only a single table (they are not joins) and have 4374 ** simple == constraints against indexed fields. This routine attempts 4375 ** to plan those simple cases using much less ceremony than the 4376 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4377 ** times for the common case. 4378 ** 4379 ** Return non-zero on success, if this query can be handled by this 4380 ** no-frills query planner. Return zero if this query needs the 4381 ** general-purpose query planner. 4382 */ 4383 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4384 WhereInfo *pWInfo; 4385 struct SrcList_item *pItem; 4386 WhereClause *pWC; 4387 WhereTerm *pTerm; 4388 WhereLoop *pLoop; 4389 int iCur; 4390 int j; 4391 Table *pTab; 4392 Index *pIdx; 4393 4394 pWInfo = pBuilder->pWInfo; 4395 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4396 assert( pWInfo->pTabList->nSrc>=1 ); 4397 pItem = pWInfo->pTabList->a; 4398 pTab = pItem->pTab; 4399 if( IsVirtual(pTab) ) return 0; 4400 if( pItem->fg.isIndexedBy ) return 0; 4401 iCur = pItem->iCursor; 4402 pWC = &pWInfo->sWC; 4403 pLoop = pBuilder->pNew; 4404 pLoop->wsFlags = 0; 4405 pLoop->nSkip = 0; 4406 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0); 4407 if( pTerm ){ 4408 testcase( pTerm->eOperator & WO_IS ); 4409 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4410 pLoop->aLTerm[0] = pTerm; 4411 pLoop->nLTerm = 1; 4412 pLoop->u.btree.nEq = 1; 4413 /* TUNING: Cost of a rowid lookup is 10 */ 4414 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 4415 }else{ 4416 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 4417 int opMask; 4418 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 4419 if( !IsUniqueIndex(pIdx) 4420 || pIdx->pPartIdxWhere!=0 4421 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 4422 ) continue; 4423 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 4424 for(j=0; j<pIdx->nKeyCol; j++){ 4425 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx); 4426 if( pTerm==0 ) break; 4427 testcase( pTerm->eOperator & WO_IS ); 4428 pLoop->aLTerm[j] = pTerm; 4429 } 4430 if( j!=pIdx->nKeyCol ) continue; 4431 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 4432 if( pIdx->isCovering || (pItem->colUsed & pIdx->colNotIdxed)==0 ){ 4433 pLoop->wsFlags |= WHERE_IDX_ONLY; 4434 } 4435 pLoop->nLTerm = j; 4436 pLoop->u.btree.nEq = j; 4437 pLoop->u.btree.pIndex = pIdx; 4438 /* TUNING: Cost of a unique index lookup is 15 */ 4439 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 4440 break; 4441 } 4442 } 4443 if( pLoop->wsFlags ){ 4444 pLoop->nOut = (LogEst)1; 4445 pWInfo->a[0].pWLoop = pLoop; 4446 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 4447 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 4448 pWInfo->a[0].iTabCur = iCur; 4449 pWInfo->nRowOut = 1; 4450 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 4451 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 4452 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4453 } 4454 #ifdef SQLITE_DEBUG 4455 pLoop->cId = '0'; 4456 #endif 4457 return 1; 4458 } 4459 return 0; 4460 } 4461 4462 /* 4463 ** Helper function for exprIsDeterministic(). 4464 */ 4465 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 4466 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 4467 pWalker->eCode = 0; 4468 return WRC_Abort; 4469 } 4470 return WRC_Continue; 4471 } 4472 4473 /* 4474 ** Return true if the expression contains no non-deterministic SQL 4475 ** functions. Do not consider non-deterministic SQL functions that are 4476 ** part of sub-select statements. 4477 */ 4478 static int exprIsDeterministic(Expr *p){ 4479 Walker w; 4480 memset(&w, 0, sizeof(w)); 4481 w.eCode = 1; 4482 w.xExprCallback = exprNodeIsDeterministic; 4483 w.xSelectCallback = sqlite3SelectWalkFail; 4484 sqlite3WalkExpr(&w, p); 4485 return w.eCode; 4486 } 4487 4488 /* 4489 ** Generate the beginning of the loop used for WHERE clause processing. 4490 ** The return value is a pointer to an opaque structure that contains 4491 ** information needed to terminate the loop. Later, the calling routine 4492 ** should invoke sqlite3WhereEnd() with the return value of this function 4493 ** in order to complete the WHERE clause processing. 4494 ** 4495 ** If an error occurs, this routine returns NULL. 4496 ** 4497 ** The basic idea is to do a nested loop, one loop for each table in 4498 ** the FROM clause of a select. (INSERT and UPDATE statements are the 4499 ** same as a SELECT with only a single table in the FROM clause.) For 4500 ** example, if the SQL is this: 4501 ** 4502 ** SELECT * FROM t1, t2, t3 WHERE ...; 4503 ** 4504 ** Then the code generated is conceptually like the following: 4505 ** 4506 ** foreach row1 in t1 do \ Code generated 4507 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 4508 ** foreach row3 in t3 do / 4509 ** ... 4510 ** end \ Code generated 4511 ** end |-- by sqlite3WhereEnd() 4512 ** end / 4513 ** 4514 ** Note that the loops might not be nested in the order in which they 4515 ** appear in the FROM clause if a different order is better able to make 4516 ** use of indices. Note also that when the IN operator appears in 4517 ** the WHERE clause, it might result in additional nested loops for 4518 ** scanning through all values on the right-hand side of the IN. 4519 ** 4520 ** There are Btree cursors associated with each table. t1 uses cursor 4521 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 4522 ** And so forth. This routine generates code to open those VDBE cursors 4523 ** and sqlite3WhereEnd() generates the code to close them. 4524 ** 4525 ** The code that sqlite3WhereBegin() generates leaves the cursors named 4526 ** in pTabList pointing at their appropriate entries. The [...] code 4527 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 4528 ** data from the various tables of the loop. 4529 ** 4530 ** If the WHERE clause is empty, the foreach loops must each scan their 4531 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 4532 ** the tables have indices and there are terms in the WHERE clause that 4533 ** refer to those indices, a complete table scan can be avoided and the 4534 ** code will run much faster. Most of the work of this routine is checking 4535 ** to see if there are indices that can be used to speed up the loop. 4536 ** 4537 ** Terms of the WHERE clause are also used to limit which rows actually 4538 ** make it to the "..." in the middle of the loop. After each "foreach", 4539 ** terms of the WHERE clause that use only terms in that loop and outer 4540 ** loops are evaluated and if false a jump is made around all subsequent 4541 ** inner loops (or around the "..." if the test occurs within the inner- 4542 ** most loop) 4543 ** 4544 ** OUTER JOINS 4545 ** 4546 ** An outer join of tables t1 and t2 is conceptally coded as follows: 4547 ** 4548 ** foreach row1 in t1 do 4549 ** flag = 0 4550 ** foreach row2 in t2 do 4551 ** start: 4552 ** ... 4553 ** flag = 1 4554 ** end 4555 ** if flag==0 then 4556 ** move the row2 cursor to a null row 4557 ** goto start 4558 ** fi 4559 ** end 4560 ** 4561 ** ORDER BY CLAUSE PROCESSING 4562 ** 4563 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 4564 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 4565 ** if there is one. If there is no ORDER BY clause or if this routine 4566 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 4567 ** 4568 ** The iIdxCur parameter is the cursor number of an index. If 4569 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 4570 ** to use for OR clause processing. The WHERE clause should use this 4571 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 4572 ** the first cursor in an array of cursors for all indices. iIdxCur should 4573 ** be used to compute the appropriate cursor depending on which index is 4574 ** used. 4575 */ 4576 WhereInfo *sqlite3WhereBegin( 4577 Parse *pParse, /* The parser context */ 4578 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 4579 Expr *pWhere, /* The WHERE clause */ 4580 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 4581 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 4582 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 4583 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 4584 ** If WHERE_USE_LIMIT, then the limit amount */ 4585 ){ 4586 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4587 int nTabList; /* Number of elements in pTabList */ 4588 WhereInfo *pWInfo; /* Will become the return value of this function */ 4589 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4590 Bitmask notReady; /* Cursors that are not yet positioned */ 4591 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 4592 WhereMaskSet *pMaskSet; /* The expression mask set */ 4593 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 4594 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 4595 int ii; /* Loop counter */ 4596 sqlite3 *db; /* Database connection */ 4597 int rc; /* Return code */ 4598 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 4599 4600 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 4601 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 4602 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4603 )); 4604 4605 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 4606 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4607 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 4608 4609 /* Variable initialization */ 4610 db = pParse->db; 4611 memset(&sWLB, 0, sizeof(sWLB)); 4612 4613 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 4614 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 4615 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 4616 sWLB.pOrderBy = pOrderBy; 4617 4618 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 4619 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 4620 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 4621 wctrlFlags &= ~WHERE_WANT_DISTINCT; 4622 } 4623 4624 /* The number of tables in the FROM clause is limited by the number of 4625 ** bits in a Bitmask 4626 */ 4627 testcase( pTabList->nSrc==BMS ); 4628 if( pTabList->nSrc>BMS ){ 4629 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 4630 return 0; 4631 } 4632 4633 /* This function normally generates a nested loop for all tables in 4634 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 4635 ** only generate code for the first table in pTabList and assume that 4636 ** any cursors associated with subsequent tables are uninitialized. 4637 */ 4638 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 4639 4640 /* Allocate and initialize the WhereInfo structure that will become the 4641 ** return value. A single allocation is used to store the WhereInfo 4642 ** struct, the contents of WhereInfo.a[], the WhereClause structure 4643 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 4644 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 4645 ** some architectures. Hence the ROUND8() below. 4646 */ 4647 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 4648 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 4649 if( db->mallocFailed ){ 4650 sqlite3DbFree(db, pWInfo); 4651 pWInfo = 0; 4652 goto whereBeginError; 4653 } 4654 pWInfo->pParse = pParse; 4655 pWInfo->pTabList = pTabList; 4656 pWInfo->pOrderBy = pOrderBy; 4657 pWInfo->pWhere = pWhere; 4658 pWInfo->pResultSet = pResultSet; 4659 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 4660 pWInfo->nLevel = nTabList; 4661 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(pParse); 4662 pWInfo->wctrlFlags = wctrlFlags; 4663 pWInfo->iLimit = iAuxArg; 4664 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 4665 memset(&pWInfo->nOBSat, 0, 4666 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 4667 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 4668 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 4669 pMaskSet = &pWInfo->sMaskSet; 4670 sWLB.pWInfo = pWInfo; 4671 sWLB.pWC = &pWInfo->sWC; 4672 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 4673 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 4674 whereLoopInit(sWLB.pNew); 4675 #ifdef SQLITE_DEBUG 4676 sWLB.pNew->cId = '*'; 4677 #endif 4678 4679 /* Split the WHERE clause into separate subexpressions where each 4680 ** subexpression is separated by an AND operator. 4681 */ 4682 initMaskSet(pMaskSet); 4683 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 4684 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 4685 4686 /* Special case: No FROM clause 4687 */ 4688 if( nTabList==0 ){ 4689 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 4690 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4691 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4692 } 4693 ExplainQueryPlan((pParse, 0, "SCAN CONSTANT ROW")); 4694 }else{ 4695 /* Assign a bit from the bitmask to every term in the FROM clause. 4696 ** 4697 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 4698 ** 4699 ** The rule of the previous sentence ensures thta if X is the bitmask for 4700 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 4701 ** Knowing the bitmask for all tables to the left of a left join is 4702 ** important. Ticket #3015. 4703 ** 4704 ** Note that bitmasks are created for all pTabList->nSrc tables in 4705 ** pTabList, not just the first nTabList tables. nTabList is normally 4706 ** equal to pTabList->nSrc but might be shortened to 1 if the 4707 ** WHERE_OR_SUBCLAUSE flag is set. 4708 */ 4709 ii = 0; 4710 do{ 4711 createMask(pMaskSet, pTabList->a[ii].iCursor); 4712 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 4713 }while( (++ii)<pTabList->nSrc ); 4714 #ifdef SQLITE_DEBUG 4715 { 4716 Bitmask mx = 0; 4717 for(ii=0; ii<pTabList->nSrc; ii++){ 4718 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 4719 assert( m>=mx ); 4720 mx = m; 4721 } 4722 } 4723 #endif 4724 } 4725 4726 /* Analyze all of the subexpressions. */ 4727 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 4728 if( db->mallocFailed ) goto whereBeginError; 4729 4730 /* Special case: WHERE terms that do not refer to any tables in the join 4731 ** (constant expressions). Evaluate each such term, and jump over all the 4732 ** generated code if the result is not true. 4733 ** 4734 ** Do not do this if the expression contains non-deterministic functions 4735 ** that are not within a sub-select. This is not strictly required, but 4736 ** preserves SQLite's legacy behaviour in the following two cases: 4737 ** 4738 ** FROM ... WHERE random()>0; -- eval random() once per row 4739 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 4740 */ 4741 for(ii=0; ii<sWLB.pWC->nTerm; ii++){ 4742 WhereTerm *pT = &sWLB.pWC->a[ii]; 4743 if( pT->wtFlags & TERM_VIRTUAL ) continue; 4744 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 4745 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 4746 pT->wtFlags |= TERM_CODED; 4747 } 4748 } 4749 4750 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4751 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 4752 /* The DISTINCT marking is pointless. Ignore it. */ 4753 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4754 }else if( pOrderBy==0 ){ 4755 /* Try to ORDER BY the result set to make distinct processing easier */ 4756 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 4757 pWInfo->pOrderBy = pResultSet; 4758 } 4759 } 4760 4761 /* Construct the WhereLoop objects */ 4762 #if defined(WHERETRACE_ENABLED) 4763 if( sqlite3WhereTrace & 0xffff ){ 4764 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 4765 if( wctrlFlags & WHERE_USE_LIMIT ){ 4766 sqlite3DebugPrintf(", limit: %d", iAuxArg); 4767 } 4768 sqlite3DebugPrintf(")\n"); 4769 } 4770 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 4771 sqlite3WhereClausePrint(sWLB.pWC); 4772 } 4773 #endif 4774 4775 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 4776 rc = whereLoopAddAll(&sWLB); 4777 if( rc ) goto whereBeginError; 4778 4779 #ifdef WHERETRACE_ENABLED 4780 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4781 WhereLoop *p; 4782 int i; 4783 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4784 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4785 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4786 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 4787 whereLoopPrint(p, sWLB.pWC); 4788 } 4789 } 4790 #endif 4791 4792 wherePathSolver(pWInfo, 0); 4793 if( db->mallocFailed ) goto whereBeginError; 4794 if( pWInfo->pOrderBy ){ 4795 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 4796 if( db->mallocFailed ) goto whereBeginError; 4797 } 4798 } 4799 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 4800 pWInfo->revMask = ALLBITS; 4801 } 4802 if( pParse->nErr || NEVER(db->mallocFailed) ){ 4803 goto whereBeginError; 4804 } 4805 #ifdef WHERETRACE_ENABLED 4806 if( sqlite3WhereTrace ){ 4807 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 4808 if( pWInfo->nOBSat>0 ){ 4809 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 4810 } 4811 switch( pWInfo->eDistinct ){ 4812 case WHERE_DISTINCT_UNIQUE: { 4813 sqlite3DebugPrintf(" DISTINCT=unique"); 4814 break; 4815 } 4816 case WHERE_DISTINCT_ORDERED: { 4817 sqlite3DebugPrintf(" DISTINCT=ordered"); 4818 break; 4819 } 4820 case WHERE_DISTINCT_UNORDERED: { 4821 sqlite3DebugPrintf(" DISTINCT=unordered"); 4822 break; 4823 } 4824 } 4825 sqlite3DebugPrintf("\n"); 4826 for(ii=0; ii<pWInfo->nLevel; ii++){ 4827 whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 4828 } 4829 } 4830 #endif 4831 4832 /* Attempt to omit tables from the join that do not affect the result. 4833 ** For a table to not affect the result, the following must be true: 4834 ** 4835 ** 1) The query must not be an aggregate. 4836 ** 2) The table must be the RHS of a LEFT JOIN. 4837 ** 3) Either the query must be DISTINCT, or else the ON or USING clause 4838 ** must contain a constraint that limits the scan of the table to 4839 ** at most a single row. 4840 ** 4) The table must not be referenced by any part of the query apart 4841 ** from its own USING or ON clause. 4842 ** 4843 ** For example, given: 4844 ** 4845 ** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1); 4846 ** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2); 4847 ** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3); 4848 ** 4849 ** then table t2 can be omitted from the following: 4850 ** 4851 ** SELECT v1, v3 FROM t1 4852 ** LEFT JOIN t2 USING (t1.ipk=t2.ipk) 4853 ** LEFT JOIN t3 USING (t1.ipk=t3.ipk) 4854 ** 4855 ** or from: 4856 ** 4857 ** SELECT DISTINCT v1, v3 FROM t1 4858 ** LEFT JOIN t2 4859 ** LEFT JOIN t3 USING (t1.ipk=t3.ipk) 4860 */ 4861 notReady = ~(Bitmask)0; 4862 if( pWInfo->nLevel>=2 4863 && pResultSet!=0 /* guarantees condition (1) above */ 4864 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 4865 ){ 4866 int i; 4867 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet); 4868 if( sWLB.pOrderBy ){ 4869 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy); 4870 } 4871 for(i=pWInfo->nLevel-1; i>=1; i--){ 4872 WhereTerm *pTerm, *pEnd; 4873 struct SrcList_item *pItem; 4874 pLoop = pWInfo->a[i].pWLoop; 4875 pItem = &pWInfo->pTabList->a[pLoop->iTab]; 4876 if( (pItem->fg.jointype & JT_LEFT)==0 ) continue; 4877 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 4878 && (pLoop->wsFlags & WHERE_ONEROW)==0 4879 ){ 4880 continue; 4881 } 4882 if( (tabUsed & pLoop->maskSelf)!=0 ) continue; 4883 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; 4884 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4885 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 4886 if( !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 4887 || pTerm->pExpr->iRightJoinTable!=pItem->iCursor 4888 ){ 4889 break; 4890 } 4891 } 4892 } 4893 if( pTerm<pEnd ) continue; 4894 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 4895 notReady &= ~pLoop->maskSelf; 4896 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4897 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 4898 pTerm->wtFlags |= TERM_CODED; 4899 } 4900 } 4901 if( i!=pWInfo->nLevel-1 ){ 4902 int nByte = (pWInfo->nLevel-1-i) * sizeof(WhereLevel); 4903 memmove(&pWInfo->a[i], &pWInfo->a[i+1], nByte); 4904 } 4905 pWInfo->nLevel--; 4906 nTabList--; 4907 } 4908 } 4909 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 4910 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 4911 4912 /* If the caller is an UPDATE or DELETE statement that is requesting 4913 ** to use a one-pass algorithm, determine if this is appropriate. 4914 ** 4915 ** A one-pass approach can be used if the caller has requested one 4916 ** and either (a) the scan visits at most one row or (b) each 4917 ** of the following are true: 4918 ** 4919 ** * the caller has indicated that a one-pass approach can be used 4920 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and 4921 ** * the table is not a virtual table, and 4922 ** * either the scan does not use the OR optimization or the caller 4923 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified 4924 ** for DELETE). 4925 ** 4926 ** The last qualification is because an UPDATE statement uses 4927 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can 4928 ** use a one-pass approach, and this is not set accurately for scans 4929 ** that use the OR optimization. 4930 */ 4931 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 4932 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 4933 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 4934 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 4935 assert( !(wsFlags & WHERE_VIRTUALTABLE) || IsVirtual(pTabList->a[0].pTab) ); 4936 if( bOnerow || ( 4937 0!=(wctrlFlags & WHERE_ONEPASS_MULTIROW) 4938 && !IsVirtual(pTabList->a[0].pTab) 4939 && (0==(wsFlags & WHERE_MULTI_OR) || (wctrlFlags & WHERE_DUPLICATES_OK)) 4940 )){ 4941 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 4942 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 4943 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 4944 bFordelete = OPFLAG_FORDELETE; 4945 } 4946 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 4947 } 4948 } 4949 } 4950 4951 /* Open all tables in the pTabList and any indices selected for 4952 ** searching those tables. 4953 */ 4954 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 4955 Table *pTab; /* Table to open */ 4956 int iDb; /* Index of database containing table/index */ 4957 struct SrcList_item *pTabItem; 4958 4959 pTabItem = &pTabList->a[pLevel->iFrom]; 4960 pTab = pTabItem->pTab; 4961 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 4962 pLoop = pLevel->pWLoop; 4963 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 4964 /* Do nothing */ 4965 }else 4966 #ifndef SQLITE_OMIT_VIRTUALTABLE 4967 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 4968 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 4969 int iCur = pTabItem->iCursor; 4970 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 4971 }else if( IsVirtual(pTab) ){ 4972 /* noop */ 4973 }else 4974 #endif 4975 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 4976 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 4977 int op = OP_OpenRead; 4978 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4979 op = OP_OpenWrite; 4980 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 4981 }; 4982 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 4983 assert( pTabItem->iCursor==pLevel->iTabCur ); 4984 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 4985 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 4986 if( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol<BMS && HasRowid(pTab) ){ 4987 Bitmask b = pTabItem->colUsed; 4988 int n = 0; 4989 for(; b; b=b>>1, n++){} 4990 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 4991 assert( n<=pTab->nCol ); 4992 } 4993 #ifdef SQLITE_ENABLE_CURSOR_HINTS 4994 if( pLoop->u.btree.pIndex!=0 ){ 4995 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 4996 }else 4997 #endif 4998 { 4999 sqlite3VdbeChangeP5(v, bFordelete); 5000 } 5001 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5002 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 5003 (const u8*)&pTabItem->colUsed, P4_INT64); 5004 #endif 5005 }else{ 5006 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 5007 } 5008 if( pLoop->wsFlags & WHERE_INDEXED ){ 5009 Index *pIx = pLoop->u.btree.pIndex; 5010 int iIndexCur; 5011 int op = OP_OpenRead; 5012 /* iAuxArg is always set to a positive value if ONEPASS is possible */ 5013 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 5014 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 5015 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 5016 ){ 5017 /* This is one term of an OR-optimization using the PRIMARY KEY of a 5018 ** WITHOUT ROWID table. No need for a separate index */ 5019 iIndexCur = pLevel->iTabCur; 5020 op = 0; 5021 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5022 Index *pJ = pTabItem->pTab->pIndex; 5023 iIndexCur = iAuxArg; 5024 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 5025 while( ALWAYS(pJ) && pJ!=pIx ){ 5026 iIndexCur++; 5027 pJ = pJ->pNext; 5028 } 5029 op = OP_OpenWrite; 5030 pWInfo->aiCurOnePass[1] = iIndexCur; 5031 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 5032 iIndexCur = iAuxArg; 5033 op = OP_ReopenIdx; 5034 }else{ 5035 iIndexCur = pParse->nTab++; 5036 } 5037 pLevel->iIdxCur = iIndexCur; 5038 assert( pIx->pSchema==pTab->pSchema ); 5039 assert( iIndexCur>=0 ); 5040 if( op ){ 5041 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 5042 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 5043 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 5044 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 5045 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 5046 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 5047 ){ 5048 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); /* Hint to COMDB2 */ 5049 } 5050 VdbeComment((v, "%s", pIx->zName)); 5051 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5052 { 5053 u64 colUsed = 0; 5054 int ii, jj; 5055 for(ii=0; ii<pIx->nColumn; ii++){ 5056 jj = pIx->aiColumn[ii]; 5057 if( jj<0 ) continue; 5058 if( jj>63 ) jj = 63; 5059 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 5060 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 5061 } 5062 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 5063 (u8*)&colUsed, P4_INT64); 5064 } 5065 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 5066 } 5067 } 5068 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 5069 } 5070 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 5071 if( db->mallocFailed ) goto whereBeginError; 5072 5073 /* Generate the code to do the search. Each iteration of the for 5074 ** loop below generates code for a single nested loop of the VM 5075 ** program. 5076 */ 5077 for(ii=0; ii<nTabList; ii++){ 5078 int addrExplain; 5079 int wsFlags; 5080 pLevel = &pWInfo->a[ii]; 5081 wsFlags = pLevel->pWLoop->wsFlags; 5082 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 5083 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 5084 constructAutomaticIndex(pParse, &pWInfo->sWC, 5085 &pTabList->a[pLevel->iFrom], notReady, pLevel); 5086 if( db->mallocFailed ) goto whereBeginError; 5087 } 5088 #endif 5089 addrExplain = sqlite3WhereExplainOneScan( 5090 pParse, pTabList, pLevel, wctrlFlags 5091 ); 5092 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 5093 notReady = sqlite3WhereCodeOneLoopStart(pParse,v,pWInfo,ii,pLevel,notReady); 5094 pWInfo->iContinue = pLevel->addrCont; 5095 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 5096 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 5097 } 5098 } 5099 5100 /* Done. */ 5101 VdbeModuleComment((v, "Begin WHERE-core")); 5102 return pWInfo; 5103 5104 /* Jump here if malloc fails */ 5105 whereBeginError: 5106 if( pWInfo ){ 5107 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5108 whereInfoFree(db, pWInfo); 5109 } 5110 return 0; 5111 } 5112 5113 /* 5114 ** Part of sqlite3WhereEnd() will rewrite opcodes to reference the 5115 ** index rather than the main table. In SQLITE_DEBUG mode, we want 5116 ** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine 5117 ** does that. 5118 */ 5119 #ifndef SQLITE_DEBUG 5120 # define OpcodeRewriteTrace(D,K,P) /* no-op */ 5121 #else 5122 # define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P) 5123 static void sqlite3WhereOpcodeRewriteTrace( 5124 sqlite3 *db, 5125 int pc, 5126 VdbeOp *pOp 5127 ){ 5128 if( (db->flags & SQLITE_VdbeAddopTrace)==0 ) return; 5129 sqlite3VdbePrintOp(0, pc, pOp); 5130 } 5131 #endif 5132 5133 /* 5134 ** Generate the end of the WHERE loop. See comments on 5135 ** sqlite3WhereBegin() for additional information. 5136 */ 5137 void sqlite3WhereEnd(WhereInfo *pWInfo){ 5138 Parse *pParse = pWInfo->pParse; 5139 Vdbe *v = pParse->pVdbe; 5140 int i; 5141 WhereLevel *pLevel; 5142 WhereLoop *pLoop; 5143 SrcList *pTabList = pWInfo->pTabList; 5144 sqlite3 *db = pParse->db; 5145 5146 /* Generate loop termination code. 5147 */ 5148 VdbeModuleComment((v, "End WHERE-core")); 5149 for(i=pWInfo->nLevel-1; i>=0; i--){ 5150 int addr; 5151 pLevel = &pWInfo->a[i]; 5152 pLoop = pLevel->pWLoop; 5153 if( pLevel->op!=OP_Noop ){ 5154 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5155 int addrSeek = 0; 5156 Index *pIdx; 5157 int n; 5158 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 5159 && i==pWInfo->nLevel-1 /* Ticket [ef9318757b152e3] 2017-10-21 */ 5160 && (pLoop->wsFlags & WHERE_INDEXED)!=0 5161 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 5162 && (n = pLoop->u.btree.nIdxCol)>0 5163 && pIdx->aiRowLogEst[n]>=36 5164 ){ 5165 int r1 = pParse->nMem+1; 5166 int j, op; 5167 for(j=0; j<n; j++){ 5168 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 5169 } 5170 pParse->nMem += n+1; 5171 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 5172 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 5173 VdbeCoverageIf(v, op==OP_SeekLT); 5174 VdbeCoverageIf(v, op==OP_SeekGT); 5175 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 5176 } 5177 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 5178 /* The common case: Advance to the next row */ 5179 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5180 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 5181 sqlite3VdbeChangeP5(v, pLevel->p5); 5182 VdbeCoverage(v); 5183 VdbeCoverageIf(v, pLevel->op==OP_Next); 5184 VdbeCoverageIf(v, pLevel->op==OP_Prev); 5185 VdbeCoverageIf(v, pLevel->op==OP_VNext); 5186 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5187 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 5188 #endif 5189 }else{ 5190 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5191 } 5192 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 5193 struct InLoop *pIn; 5194 int j; 5195 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 5196 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 5197 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5198 if( pIn->eEndLoopOp!=OP_Noop ){ 5199 if( pIn->nPrefix ){ 5200 assert( pLoop->wsFlags & WHERE_IN_EARLYOUT ); 5201 sqlite3VdbeAddOp4Int(v, OP_IfNoHope, pLevel->iIdxCur, 5202 sqlite3VdbeCurrentAddr(v)+2, 5203 pIn->iBase, pIn->nPrefix); 5204 VdbeCoverage(v); 5205 } 5206 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 5207 VdbeCoverage(v); 5208 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Prev); 5209 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Next); 5210 } 5211 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 5212 } 5213 } 5214 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 5215 if( pLevel->addrSkip ){ 5216 sqlite3VdbeGoto(v, pLevel->addrSkip); 5217 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 5218 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 5219 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 5220 } 5221 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 5222 if( pLevel->addrLikeRep ){ 5223 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 5224 pLevel->addrLikeRep); 5225 VdbeCoverage(v); 5226 } 5227 #endif 5228 if( pLevel->iLeftJoin ){ 5229 int ws = pLoop->wsFlags; 5230 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 5231 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 5232 if( (ws & WHERE_IDX_ONLY)==0 ){ 5233 assert( pLevel->iTabCur==pTabList->a[pLevel->iFrom].iCursor ); 5234 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iTabCur); 5235 } 5236 if( (ws & WHERE_INDEXED) 5237 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCovidx) 5238 ){ 5239 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 5240 } 5241 if( pLevel->op==OP_Return ){ 5242 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 5243 }else{ 5244 sqlite3VdbeGoto(v, pLevel->addrFirst); 5245 } 5246 sqlite3VdbeJumpHere(v, addr); 5247 } 5248 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 5249 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 5250 } 5251 5252 /* The "break" point is here, just past the end of the outer loop. 5253 ** Set it. 5254 */ 5255 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5256 5257 assert( pWInfo->nLevel<=pTabList->nSrc ); 5258 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5259 int k, last; 5260 VdbeOp *pOp; 5261 Index *pIdx = 0; 5262 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; 5263 Table *pTab = pTabItem->pTab; 5264 assert( pTab!=0 ); 5265 pLoop = pLevel->pWLoop; 5266 5267 /* For a co-routine, change all OP_Column references to the table of 5268 ** the co-routine into OP_Copy of result contained in a register. 5269 ** OP_Rowid becomes OP_Null. 5270 */ 5271 if( pTabItem->fg.viaCoroutine ){ 5272 testcase( pParse->db->mallocFailed ); 5273 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 5274 pTabItem->regResult, 0); 5275 continue; 5276 } 5277 5278 #ifdef SQLITE_ENABLE_EARLY_CURSOR_CLOSE 5279 /* Close all of the cursors that were opened by sqlite3WhereBegin. 5280 ** Except, do not close cursors that will be reused by the OR optimization 5281 ** (WHERE_OR_SUBCLAUSE). And do not close the OP_OpenWrite cursors 5282 ** created for the ONEPASS optimization. 5283 */ 5284 if( (pTab->tabFlags & TF_Ephemeral)==0 5285 && pTab->pSelect==0 5286 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5287 ){ 5288 int ws = pLoop->wsFlags; 5289 if( pWInfo->eOnePass==ONEPASS_OFF && (ws & WHERE_IDX_ONLY)==0 ){ 5290 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); 5291 } 5292 if( (ws & WHERE_INDEXED)!=0 5293 && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0 5294 && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1] 5295 ){ 5296 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); 5297 } 5298 } 5299 #endif 5300 5301 /* If this scan uses an index, make VDBE code substitutions to read data 5302 ** from the index instead of from the table where possible. In some cases 5303 ** this optimization prevents the table from ever being read, which can 5304 ** yield a significant performance boost. 5305 ** 5306 ** Calls to the code generator in between sqlite3WhereBegin and 5307 ** sqlite3WhereEnd will have created code that references the table 5308 ** directly. This loop scans all that code looking for opcodes 5309 ** that reference the table and converts them into opcodes that 5310 ** reference the index. 5311 */ 5312 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 5313 pIdx = pLoop->u.btree.pIndex; 5314 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 5315 pIdx = pLevel->u.pCovidx; 5316 } 5317 if( pIdx 5318 && (pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable)) 5319 && !db->mallocFailed 5320 ){ 5321 last = sqlite3VdbeCurrentAddr(v); 5322 k = pLevel->addrBody; 5323 #ifdef SQLITE_DEBUG 5324 if( db->flags & SQLITE_VdbeAddopTrace ){ 5325 printf("TRANSLATE opcodes in range %d..%d\n", k, last-1); 5326 } 5327 #endif 5328 pOp = sqlite3VdbeGetOp(v, k); 5329 for(; k<last; k++, pOp++){ 5330 if( pOp->p1!=pLevel->iTabCur ) continue; 5331 if( pOp->opcode==OP_Column 5332 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC 5333 || pOp->opcode==OP_Offset 5334 #endif 5335 ){ 5336 int x = pOp->p2; 5337 assert( pIdx->pTable==pTab ); 5338 if( !HasRowid(pTab) ){ 5339 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 5340 x = pPk->aiColumn[x]; 5341 assert( x>=0 ); 5342 } 5343 x = sqlite3ColumnOfIndex(pIdx, x); 5344 if( x>=0 ){ 5345 pOp->p2 = x; 5346 pOp->p1 = pLevel->iIdxCur; 5347 OpcodeRewriteTrace(db, k, pOp); 5348 } 5349 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 5350 || pWInfo->eOnePass ); 5351 }else if( pOp->opcode==OP_Rowid ){ 5352 pOp->p1 = pLevel->iIdxCur; 5353 pOp->opcode = OP_IdxRowid; 5354 OpcodeRewriteTrace(db, k, pOp); 5355 }else if( pOp->opcode==OP_IfNullRow ){ 5356 pOp->p1 = pLevel->iIdxCur; 5357 OpcodeRewriteTrace(db, k, pOp); 5358 } 5359 } 5360 #ifdef SQLITE_DEBUG 5361 if( db->flags & SQLITE_VdbeAddopTrace ) printf("TRANSLATE complete\n"); 5362 #endif 5363 } 5364 } 5365 5366 /* Final cleanup 5367 */ 5368 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5369 whereInfoFree(db, pWInfo); 5370 return; 5371 } 5372