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_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_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_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 assert( pIdx->zColAff[iCol]!=0 ); 1309 return pIdx->zColAff[iCol]; 1310 } 1311 #endif 1312 1313 1314 #ifdef SQLITE_ENABLE_STAT4 1315 /* 1316 ** This function is called to estimate the number of rows visited by a 1317 ** range-scan on a skip-scan index. For example: 1318 ** 1319 ** CREATE INDEX i1 ON t1(a, b, c); 1320 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1321 ** 1322 ** Value pLoop->nOut is currently set to the estimated number of rows 1323 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1324 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1325 ** on the stat4 data for the index. this scan will be peformed multiple 1326 ** times (once for each (a,b) combination that matches a=?) is dealt with 1327 ** by the caller. 1328 ** 1329 ** It does this by scanning through all stat4 samples, comparing values 1330 ** extracted from pLower and pUpper with the corresponding column in each 1331 ** sample. If L and U are the number of samples found to be less than or 1332 ** equal to the values extracted from pLower and pUpper respectively, and 1333 ** N is the total number of samples, the pLoop->nOut value is adjusted 1334 ** as follows: 1335 ** 1336 ** nOut = nOut * ( min(U - L, 1) / N ) 1337 ** 1338 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1339 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1340 ** U is set to N. 1341 ** 1342 ** Normally, this function sets *pbDone to 1 before returning. However, 1343 ** if no value can be extracted from either pLower or pUpper (and so the 1344 ** estimate of the number of rows delivered remains unchanged), *pbDone 1345 ** is left as is. 1346 ** 1347 ** If an error occurs, an SQLite error code is returned. Otherwise, 1348 ** SQLITE_OK. 1349 */ 1350 static int whereRangeSkipScanEst( 1351 Parse *pParse, /* Parsing & code generating context */ 1352 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1353 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1354 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1355 int *pbDone /* Set to true if at least one expr. value extracted */ 1356 ){ 1357 Index *p = pLoop->u.btree.pIndex; 1358 int nEq = pLoop->u.btree.nEq; 1359 sqlite3 *db = pParse->db; 1360 int nLower = -1; 1361 int nUpper = p->nSample+1; 1362 int rc = SQLITE_OK; 1363 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1364 CollSeq *pColl; 1365 1366 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1367 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1368 sqlite3_value *pVal = 0; /* Value extracted from record */ 1369 1370 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1371 if( pLower ){ 1372 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1373 nLower = 0; 1374 } 1375 if( pUpper && rc==SQLITE_OK ){ 1376 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1377 nUpper = p2 ? 0 : p->nSample; 1378 } 1379 1380 if( p1 || p2 ){ 1381 int i; 1382 int nDiff; 1383 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1384 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1385 if( rc==SQLITE_OK && p1 ){ 1386 int res = sqlite3MemCompare(p1, pVal, pColl); 1387 if( res>=0 ) nLower++; 1388 } 1389 if( rc==SQLITE_OK && p2 ){ 1390 int res = sqlite3MemCompare(p2, pVal, pColl); 1391 if( res>=0 ) nUpper++; 1392 } 1393 } 1394 nDiff = (nUpper - nLower); 1395 if( nDiff<=0 ) nDiff = 1; 1396 1397 /* If there is both an upper and lower bound specified, and the 1398 ** comparisons indicate that they are close together, use the fallback 1399 ** method (assume that the scan visits 1/64 of the rows) for estimating 1400 ** the number of rows visited. Otherwise, estimate the number of rows 1401 ** using the method described in the header comment for this function. */ 1402 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1403 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1404 pLoop->nOut -= nAdjust; 1405 *pbDone = 1; 1406 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1407 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1408 } 1409 1410 }else{ 1411 assert( *pbDone==0 ); 1412 } 1413 1414 sqlite3ValueFree(p1); 1415 sqlite3ValueFree(p2); 1416 sqlite3ValueFree(pVal); 1417 1418 return rc; 1419 } 1420 #endif /* SQLITE_ENABLE_STAT4 */ 1421 1422 /* 1423 ** This function is used to estimate the number of rows that will be visited 1424 ** by scanning an index for a range of values. The range may have an upper 1425 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1426 ** and lower bounds are represented by pLower and pUpper respectively. For 1427 ** example, assuming that index p is on t1(a): 1428 ** 1429 ** ... FROM t1 WHERE a > ? AND a < ? ... 1430 ** |_____| |_____| 1431 ** | | 1432 ** pLower pUpper 1433 ** 1434 ** If either of the upper or lower bound is not present, then NULL is passed in 1435 ** place of the corresponding WhereTerm. 1436 ** 1437 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1438 ** column subject to the range constraint. Or, equivalently, the number of 1439 ** equality constraints optimized by the proposed index scan. For example, 1440 ** assuming index p is on t1(a, b), and the SQL query is: 1441 ** 1442 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1443 ** 1444 ** then nEq is set to 1 (as the range restricted column, b, is the second 1445 ** left-most column of the index). Or, if the query is: 1446 ** 1447 ** ... FROM t1 WHERE a > ? AND a < ? ... 1448 ** 1449 ** then nEq is set to 0. 1450 ** 1451 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1452 ** number of rows that the index scan is expected to visit without 1453 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1454 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1455 ** to account for the range constraints pLower and pUpper. 1456 ** 1457 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1458 ** used, a single range inequality reduces the search space by a factor of 4. 1459 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1460 ** rows visited by a factor of 64. 1461 */ 1462 static int whereRangeScanEst( 1463 Parse *pParse, /* Parsing & code generating context */ 1464 WhereLoopBuilder *pBuilder, 1465 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1466 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1467 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1468 ){ 1469 int rc = SQLITE_OK; 1470 int nOut = pLoop->nOut; 1471 LogEst nNew; 1472 1473 #ifdef SQLITE_ENABLE_STAT4 1474 Index *p = pLoop->u.btree.pIndex; 1475 int nEq = pLoop->u.btree.nEq; 1476 1477 if( p->nSample>0 && ALWAYS(nEq<p->nSampleCol) 1478 && OptimizationEnabled(pParse->db, SQLITE_Stat4) 1479 ){ 1480 if( nEq==pBuilder->nRecValid ){ 1481 UnpackedRecord *pRec = pBuilder->pRec; 1482 tRowcnt a[2]; 1483 int nBtm = pLoop->u.btree.nBtm; 1484 int nTop = pLoop->u.btree.nTop; 1485 1486 /* Variable iLower will be set to the estimate of the number of rows in 1487 ** the index that are less than the lower bound of the range query. The 1488 ** lower bound being the concatenation of $P and $L, where $P is the 1489 ** key-prefix formed by the nEq values matched against the nEq left-most 1490 ** columns of the index, and $L is the value in pLower. 1491 ** 1492 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1493 ** is not a simple variable or literal value), the lower bound of the 1494 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1495 ** if $L is available, whereKeyStats() is called for both ($P) and 1496 ** ($P:$L) and the larger of the two returned values is used. 1497 ** 1498 ** Similarly, iUpper is to be set to the estimate of the number of rows 1499 ** less than the upper bound of the range query. Where the upper bound 1500 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1501 ** of iUpper are requested of whereKeyStats() and the smaller used. 1502 ** 1503 ** The number of rows between the two bounds is then just iUpper-iLower. 1504 */ 1505 tRowcnt iLower; /* Rows less than the lower bound */ 1506 tRowcnt iUpper; /* Rows less than the upper bound */ 1507 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1508 int iUprIdx = -1; /* aSample[] for the upper bound */ 1509 1510 if( pRec ){ 1511 testcase( pRec->nField!=pBuilder->nRecValid ); 1512 pRec->nField = pBuilder->nRecValid; 1513 } 1514 /* Determine iLower and iUpper using ($P) only. */ 1515 if( nEq==0 ){ 1516 iLower = 0; 1517 iUpper = p->nRowEst0; 1518 }else{ 1519 /* Note: this call could be optimized away - since the same values must 1520 ** have been requested when testing key $P in whereEqualScanEst(). */ 1521 whereKeyStats(pParse, p, pRec, 0, a); 1522 iLower = a[0]; 1523 iUpper = a[0] + a[1]; 1524 } 1525 1526 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1527 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1528 assert( p->aSortOrder!=0 ); 1529 if( p->aSortOrder[nEq] ){ 1530 /* The roles of pLower and pUpper are swapped for a DESC index */ 1531 SWAP(WhereTerm*, pLower, pUpper); 1532 SWAP(int, nBtm, nTop); 1533 } 1534 1535 /* If possible, improve on the iLower estimate using ($P:$L). */ 1536 if( pLower ){ 1537 int n; /* Values extracted from pExpr */ 1538 Expr *pExpr = pLower->pExpr->pRight; 1539 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1540 if( rc==SQLITE_OK && n ){ 1541 tRowcnt iNew; 1542 u16 mask = WO_GT|WO_LE; 1543 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1544 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1545 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1546 if( iNew>iLower ) iLower = iNew; 1547 nOut--; 1548 pLower = 0; 1549 } 1550 } 1551 1552 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1553 if( pUpper ){ 1554 int n; /* Values extracted from pExpr */ 1555 Expr *pExpr = pUpper->pExpr->pRight; 1556 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1557 if( rc==SQLITE_OK && n ){ 1558 tRowcnt iNew; 1559 u16 mask = WO_GT|WO_LE; 1560 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1561 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1562 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1563 if( iNew<iUpper ) iUpper = iNew; 1564 nOut--; 1565 pUpper = 0; 1566 } 1567 } 1568 1569 pBuilder->pRec = pRec; 1570 if( rc==SQLITE_OK ){ 1571 if( iUpper>iLower ){ 1572 nNew = sqlite3LogEst(iUpper - iLower); 1573 /* TUNING: If both iUpper and iLower are derived from the same 1574 ** sample, then assume they are 4x more selective. This brings 1575 ** the estimated selectivity more in line with what it would be 1576 ** if estimated without the use of STAT4 tables. */ 1577 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1578 }else{ 1579 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1580 } 1581 if( nNew<nOut ){ 1582 nOut = nNew; 1583 } 1584 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1585 (u32)iLower, (u32)iUpper, nOut)); 1586 } 1587 }else{ 1588 int bDone = 0; 1589 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1590 if( bDone ) return rc; 1591 } 1592 } 1593 #else 1594 UNUSED_PARAMETER(pParse); 1595 UNUSED_PARAMETER(pBuilder); 1596 assert( pLower || pUpper ); 1597 #endif 1598 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1599 nNew = whereRangeAdjust(pLower, nOut); 1600 nNew = whereRangeAdjust(pUpper, nNew); 1601 1602 /* TUNING: If there is both an upper and lower limit and neither limit 1603 ** has an application-defined likelihood(), assume the range is 1604 ** reduced by an additional 75%. This means that, by default, an open-ended 1605 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1606 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1607 ** match 1/64 of the index. */ 1608 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1609 nNew -= 20; 1610 } 1611 1612 nOut -= (pLower!=0) + (pUpper!=0); 1613 if( nNew<10 ) nNew = 10; 1614 if( nNew<nOut ) nOut = nNew; 1615 #if defined(WHERETRACE_ENABLED) 1616 if( pLoop->nOut>nOut ){ 1617 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1618 pLoop->nOut, nOut)); 1619 } 1620 #endif 1621 pLoop->nOut = (LogEst)nOut; 1622 return rc; 1623 } 1624 1625 #ifdef SQLITE_ENABLE_STAT4 1626 /* 1627 ** Estimate the number of rows that will be returned based on 1628 ** an equality constraint x=VALUE and where that VALUE occurs in 1629 ** the histogram data. This only works when x is the left-most 1630 ** column of an index and sqlite_stat4 histogram data is available 1631 ** for that index. When pExpr==NULL that means the constraint is 1632 ** "x IS NULL" instead of "x=VALUE". 1633 ** 1634 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1635 ** If unable to make an estimate, leave *pnRow unchanged and return 1636 ** non-zero. 1637 ** 1638 ** This routine can fail if it is unable to load a collating sequence 1639 ** required for string comparison, or if unable to allocate memory 1640 ** for a UTF conversion required for comparison. The error is stored 1641 ** in the pParse structure. 1642 */ 1643 static int whereEqualScanEst( 1644 Parse *pParse, /* Parsing & code generating context */ 1645 WhereLoopBuilder *pBuilder, 1646 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1647 tRowcnt *pnRow /* Write the revised row estimate here */ 1648 ){ 1649 Index *p = pBuilder->pNew->u.btree.pIndex; 1650 int nEq = pBuilder->pNew->u.btree.nEq; 1651 UnpackedRecord *pRec = pBuilder->pRec; 1652 int rc; /* Subfunction return code */ 1653 tRowcnt a[2]; /* Statistics */ 1654 int bOk; 1655 1656 assert( nEq>=1 ); 1657 assert( nEq<=p->nColumn ); 1658 assert( p->aSample!=0 ); 1659 assert( p->nSample>0 ); 1660 assert( pBuilder->nRecValid<nEq ); 1661 1662 /* If values are not available for all fields of the index to the left 1663 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1664 if( pBuilder->nRecValid<(nEq-1) ){ 1665 return SQLITE_NOTFOUND; 1666 } 1667 1668 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1669 ** below would return the same value. */ 1670 if( nEq>=p->nColumn ){ 1671 *pnRow = 1; 1672 return SQLITE_OK; 1673 } 1674 1675 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1676 pBuilder->pRec = pRec; 1677 if( rc!=SQLITE_OK ) return rc; 1678 if( bOk==0 ) return SQLITE_NOTFOUND; 1679 pBuilder->nRecValid = nEq; 1680 1681 whereKeyStats(pParse, p, pRec, 0, a); 1682 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1683 p->zName, nEq-1, (int)a[1])); 1684 *pnRow = a[1]; 1685 1686 return rc; 1687 } 1688 #endif /* SQLITE_ENABLE_STAT4 */ 1689 1690 #ifdef SQLITE_ENABLE_STAT4 1691 /* 1692 ** Estimate the number of rows that will be returned based on 1693 ** an IN constraint where the right-hand side of the IN operator 1694 ** is a list of values. Example: 1695 ** 1696 ** WHERE x IN (1,2,3,4) 1697 ** 1698 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1699 ** If unable to make an estimate, leave *pnRow unchanged and return 1700 ** non-zero. 1701 ** 1702 ** This routine can fail if it is unable to load a collating sequence 1703 ** required for string comparison, or if unable to allocate memory 1704 ** for a UTF conversion required for comparison. The error is stored 1705 ** in the pParse structure. 1706 */ 1707 static int whereInScanEst( 1708 Parse *pParse, /* Parsing & code generating context */ 1709 WhereLoopBuilder *pBuilder, 1710 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1711 tRowcnt *pnRow /* Write the revised row estimate here */ 1712 ){ 1713 Index *p = pBuilder->pNew->u.btree.pIndex; 1714 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1715 int nRecValid = pBuilder->nRecValid; 1716 int rc = SQLITE_OK; /* Subfunction return code */ 1717 tRowcnt nEst; /* Number of rows for a single term */ 1718 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1719 int i; /* Loop counter */ 1720 1721 assert( p->aSample!=0 ); 1722 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1723 nEst = nRow0; 1724 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1725 nRowEst += nEst; 1726 pBuilder->nRecValid = nRecValid; 1727 } 1728 1729 if( rc==SQLITE_OK ){ 1730 if( nRowEst > nRow0 ) nRowEst = nRow0; 1731 *pnRow = nRowEst; 1732 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1733 } 1734 assert( pBuilder->nRecValid==nRecValid ); 1735 return rc; 1736 } 1737 #endif /* SQLITE_ENABLE_STAT4 */ 1738 1739 1740 #ifdef WHERETRACE_ENABLED 1741 /* 1742 ** Print the content of a WhereTerm object 1743 */ 1744 static void whereTermPrint(WhereTerm *pTerm, int iTerm){ 1745 if( pTerm==0 ){ 1746 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1747 }else{ 1748 char zType[4]; 1749 char zLeft[50]; 1750 memcpy(zType, "...", 4); 1751 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1752 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1753 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1754 if( pTerm->eOperator & WO_SINGLE ){ 1755 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 1756 pTerm->leftCursor, pTerm->u.leftColumn); 1757 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 1758 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%lld", 1759 pTerm->u.pOrInfo->indexable); 1760 }else{ 1761 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 1762 } 1763 sqlite3DebugPrintf( 1764 "TERM-%-3d %p %s %-12s prob=%-3d op=0x%03x wtFlags=0x%04x", 1765 iTerm, pTerm, zType, zLeft, pTerm->truthProb, 1766 pTerm->eOperator, pTerm->wtFlags); 1767 if( pTerm->iField ){ 1768 sqlite3DebugPrintf(" iField=%d\n", pTerm->iField); 1769 }else{ 1770 sqlite3DebugPrintf("\n"); 1771 } 1772 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 1773 } 1774 } 1775 #endif 1776 1777 #ifdef WHERETRACE_ENABLED 1778 /* 1779 ** Show the complete content of a WhereClause 1780 */ 1781 void sqlite3WhereClausePrint(WhereClause *pWC){ 1782 int i; 1783 for(i=0; i<pWC->nTerm; i++){ 1784 whereTermPrint(&pWC->a[i], i); 1785 } 1786 } 1787 #endif 1788 1789 #ifdef WHERETRACE_ENABLED 1790 /* 1791 ** Print a WhereLoop object for debugging purposes 1792 */ 1793 static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){ 1794 WhereInfo *pWInfo = pWC->pWInfo; 1795 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 1796 struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab; 1797 Table *pTab = pItem->pTab; 1798 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 1799 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 1800 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 1801 sqlite3DebugPrintf(" %12s", 1802 pItem->zAlias ? pItem->zAlias : pTab->zName); 1803 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 1804 const char *zName; 1805 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 1806 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 1807 int i = sqlite3Strlen30(zName) - 1; 1808 while( zName[i]!='_' ) i--; 1809 zName += i; 1810 } 1811 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 1812 }else{ 1813 sqlite3DebugPrintf("%20s",""); 1814 } 1815 }else{ 1816 char *z; 1817 if( p->u.vtab.idxStr ){ 1818 z = sqlite3_mprintf("(%d,\"%s\",%x)", 1819 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 1820 }else{ 1821 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 1822 } 1823 sqlite3DebugPrintf(" %-19s", z); 1824 sqlite3_free(z); 1825 } 1826 if( p->wsFlags & WHERE_SKIPSCAN ){ 1827 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 1828 }else{ 1829 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); 1830 } 1831 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 1832 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 1833 int i; 1834 for(i=0; i<p->nLTerm; i++){ 1835 whereTermPrint(p->aLTerm[i], i); 1836 } 1837 } 1838 } 1839 #endif 1840 1841 /* 1842 ** Convert bulk memory into a valid WhereLoop that can be passed 1843 ** to whereLoopClear harmlessly. 1844 */ 1845 static void whereLoopInit(WhereLoop *p){ 1846 p->aLTerm = p->aLTermSpace; 1847 p->nLTerm = 0; 1848 p->nLSlot = ArraySize(p->aLTermSpace); 1849 p->wsFlags = 0; 1850 } 1851 1852 /* 1853 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 1854 */ 1855 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 1856 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 1857 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 1858 sqlite3_free(p->u.vtab.idxStr); 1859 p->u.vtab.needFree = 0; 1860 p->u.vtab.idxStr = 0; 1861 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 1862 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 1863 sqlite3DbFreeNN(db, p->u.btree.pIndex); 1864 p->u.btree.pIndex = 0; 1865 } 1866 } 1867 } 1868 1869 /* 1870 ** Deallocate internal memory used by a WhereLoop object 1871 */ 1872 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 1873 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1874 whereLoopClearUnion(db, p); 1875 whereLoopInit(p); 1876 } 1877 1878 /* 1879 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 1880 */ 1881 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 1882 WhereTerm **paNew; 1883 if( p->nLSlot>=n ) return SQLITE_OK; 1884 n = (n+7)&~7; 1885 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 1886 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 1887 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 1888 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1889 p->aLTerm = paNew; 1890 p->nLSlot = n; 1891 return SQLITE_OK; 1892 } 1893 1894 /* 1895 ** Transfer content from the second pLoop into the first. 1896 */ 1897 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 1898 whereLoopClearUnion(db, pTo); 1899 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 1900 memset(&pTo->u, 0, sizeof(pTo->u)); 1901 return SQLITE_NOMEM_BKPT; 1902 } 1903 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 1904 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 1905 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 1906 pFrom->u.vtab.needFree = 0; 1907 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 1908 pFrom->u.btree.pIndex = 0; 1909 } 1910 return SQLITE_OK; 1911 } 1912 1913 /* 1914 ** Delete a WhereLoop object 1915 */ 1916 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 1917 whereLoopClear(db, p); 1918 sqlite3DbFreeNN(db, p); 1919 } 1920 1921 /* 1922 ** Free a WhereInfo structure 1923 */ 1924 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 1925 int i; 1926 assert( pWInfo!=0 ); 1927 for(i=0; i<pWInfo->nLevel; i++){ 1928 WhereLevel *pLevel = &pWInfo->a[i]; 1929 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){ 1930 sqlite3DbFree(db, pLevel->u.in.aInLoop); 1931 } 1932 } 1933 sqlite3WhereClauseClear(&pWInfo->sWC); 1934 while( pWInfo->pLoops ){ 1935 WhereLoop *p = pWInfo->pLoops; 1936 pWInfo->pLoops = p->pNextLoop; 1937 whereLoopDelete(db, p); 1938 } 1939 sqlite3DbFreeNN(db, pWInfo); 1940 } 1941 1942 /* 1943 ** Return TRUE if all of the following are true: 1944 ** 1945 ** (1) X has the same or lower cost that Y 1946 ** (2) X uses fewer WHERE clause terms than Y 1947 ** (3) Every WHERE clause term used by X is also used by Y 1948 ** (4) X skips at least as many columns as Y 1949 ** (5) If X is a covering index, than Y is too 1950 ** 1951 ** Conditions (2) and (3) mean that X is a "proper subset" of Y. 1952 ** If X is a proper subset of Y then Y is a better choice and ought 1953 ** to have a lower cost. This routine returns TRUE when that cost 1954 ** relationship is inverted and needs to be adjusted. Constraint (4) 1955 ** was added because if X uses skip-scan less than Y it still might 1956 ** deserve a lower cost even if it is a proper subset of Y. Constraint (5) 1957 ** was added because a covering index probably deserves to have a lower cost 1958 ** than a non-covering index even if it is a proper subset. 1959 */ 1960 static int whereLoopCheaperProperSubset( 1961 const WhereLoop *pX, /* First WhereLoop to compare */ 1962 const WhereLoop *pY /* Compare against this WhereLoop */ 1963 ){ 1964 int i, j; 1965 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 1966 return 0; /* X is not a subset of Y */ 1967 } 1968 if( pY->nSkip > pX->nSkip ) return 0; 1969 if( pX->rRun >= pY->rRun ){ 1970 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ 1971 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ 1972 } 1973 for(i=pX->nLTerm-1; i>=0; i--){ 1974 if( pX->aLTerm[i]==0 ) continue; 1975 for(j=pY->nLTerm-1; j>=0; j--){ 1976 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 1977 } 1978 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 1979 } 1980 if( (pX->wsFlags&WHERE_IDX_ONLY)!=0 1981 && (pY->wsFlags&WHERE_IDX_ONLY)==0 ){ 1982 return 0; /* Constraint (5) */ 1983 } 1984 return 1; /* All conditions meet */ 1985 } 1986 1987 /* 1988 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so 1989 ** that: 1990 ** 1991 ** (1) pTemplate costs less than any other WhereLoops that are a proper 1992 ** subset of pTemplate 1993 ** 1994 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 1995 ** is a proper subset. 1996 ** 1997 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 1998 ** WHERE clause terms than Y and that every WHERE clause term used by X is 1999 ** also used by Y. 2000 */ 2001 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 2002 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 2003 for(; p; p=p->pNextLoop){ 2004 if( p->iTab!=pTemplate->iTab ) continue; 2005 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 2006 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 2007 /* Adjust pTemplate cost downward so that it is cheaper than its 2008 ** subset p. */ 2009 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2010 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); 2011 pTemplate->rRun = p->rRun; 2012 pTemplate->nOut = p->nOut - 1; 2013 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 2014 /* Adjust pTemplate cost upward so that it is costlier than p since 2015 ** pTemplate is a proper subset of p */ 2016 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2017 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); 2018 pTemplate->rRun = p->rRun; 2019 pTemplate->nOut = p->nOut + 1; 2020 } 2021 } 2022 } 2023 2024 /* 2025 ** Search the list of WhereLoops in *ppPrev looking for one that can be 2026 ** replaced by pTemplate. 2027 ** 2028 ** Return NULL if pTemplate does not belong on the WhereLoop list. 2029 ** In other words if pTemplate ought to be dropped from further consideration. 2030 ** 2031 ** If pX is a WhereLoop that pTemplate can replace, then return the 2032 ** link that points to pX. 2033 ** 2034 ** If pTemplate cannot replace any existing element of the list but needs 2035 ** to be added to the list as a new entry, then return a pointer to the 2036 ** tail of the list. 2037 */ 2038 static WhereLoop **whereLoopFindLesser( 2039 WhereLoop **ppPrev, 2040 const WhereLoop *pTemplate 2041 ){ 2042 WhereLoop *p; 2043 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 2044 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 2045 /* If either the iTab or iSortIdx values for two WhereLoop are different 2046 ** then those WhereLoops need to be considered separately. Neither is 2047 ** a candidate to replace the other. */ 2048 continue; 2049 } 2050 /* In the current implementation, the rSetup value is either zero 2051 ** or the cost of building an automatic index (NlogN) and the NlogN 2052 ** is the same for compatible WhereLoops. */ 2053 assert( p->rSetup==0 || pTemplate->rSetup==0 2054 || p->rSetup==pTemplate->rSetup ); 2055 2056 /* whereLoopAddBtree() always generates and inserts the automatic index 2057 ** case first. Hence compatible candidate WhereLoops never have a larger 2058 ** rSetup. Call this SETUP-INVARIANT */ 2059 assert( p->rSetup>=pTemplate->rSetup ); 2060 2061 /* Any loop using an appliation-defined index (or PRIMARY KEY or 2062 ** UNIQUE constraint) with one or more == constraints is better 2063 ** than an automatic index. Unless it is a skip-scan. */ 2064 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 2065 && (pTemplate->nSkip)==0 2066 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 2067 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 2068 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2069 ){ 2070 break; 2071 } 2072 2073 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2074 ** discarded. WhereLoop p is better if: 2075 ** (1) p has no more dependencies than pTemplate, and 2076 ** (2) p has an equal or lower cost than pTemplate 2077 */ 2078 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2079 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2080 && p->rRun<=pTemplate->rRun /* (2b) */ 2081 && p->nOut<=pTemplate->nOut /* (2c) */ 2082 ){ 2083 return 0; /* Discard pTemplate */ 2084 } 2085 2086 /* If pTemplate is always better than p, then cause p to be overwritten 2087 ** with pTemplate. pTemplate is better than p if: 2088 ** (1) pTemplate has no more dependences than p, and 2089 ** (2) pTemplate has an equal or lower cost than p. 2090 */ 2091 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2092 && p->rRun>=pTemplate->rRun /* (2a) */ 2093 && p->nOut>=pTemplate->nOut /* (2b) */ 2094 ){ 2095 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2096 break; /* Cause p to be overwritten by pTemplate */ 2097 } 2098 } 2099 return ppPrev; 2100 } 2101 2102 /* 2103 ** Insert or replace a WhereLoop entry using the template supplied. 2104 ** 2105 ** An existing WhereLoop entry might be overwritten if the new template 2106 ** is better and has fewer dependencies. Or the template will be ignored 2107 ** and no insert will occur if an existing WhereLoop is faster and has 2108 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2109 ** added based on the template. 2110 ** 2111 ** If pBuilder->pOrSet is not NULL then we care about only the 2112 ** prerequisites and rRun and nOut costs of the N best loops. That 2113 ** information is gathered in the pBuilder->pOrSet object. This special 2114 ** processing mode is used only for OR clause processing. 2115 ** 2116 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2117 ** still might overwrite similar loops with the new template if the 2118 ** new template is better. Loops may be overwritten if the following 2119 ** conditions are met: 2120 ** 2121 ** (1) They have the same iTab. 2122 ** (2) They have the same iSortIdx. 2123 ** (3) The template has same or fewer dependencies than the current loop 2124 ** (4) The template has the same or lower cost than the current loop 2125 */ 2126 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2127 WhereLoop **ppPrev, *p; 2128 WhereInfo *pWInfo = pBuilder->pWInfo; 2129 sqlite3 *db = pWInfo->pParse->db; 2130 int rc; 2131 2132 /* Stop the search once we hit the query planner search limit */ 2133 if( pBuilder->iPlanLimit==0 ){ 2134 WHERETRACE(0xffffffff,("=== query planner search limit reached ===\n")); 2135 if( pBuilder->pOrSet ) pBuilder->pOrSet->n = 0; 2136 return SQLITE_DONE; 2137 } 2138 pBuilder->iPlanLimit--; 2139 2140 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2141 ** and prereqs. 2142 */ 2143 if( pBuilder->pOrSet!=0 ){ 2144 if( pTemplate->nLTerm ){ 2145 #if WHERETRACE_ENABLED 2146 u16 n = pBuilder->pOrSet->n; 2147 int x = 2148 #endif 2149 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2150 pTemplate->nOut); 2151 #if WHERETRACE_ENABLED /* 0x8 */ 2152 if( sqlite3WhereTrace & 0x8 ){ 2153 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2154 whereLoopPrint(pTemplate, pBuilder->pWC); 2155 } 2156 #endif 2157 } 2158 return SQLITE_OK; 2159 } 2160 2161 /* Look for an existing WhereLoop to replace with pTemplate 2162 */ 2163 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2164 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2165 2166 if( ppPrev==0 ){ 2167 /* There already exists a WhereLoop on the list that is better 2168 ** than pTemplate, so just ignore pTemplate */ 2169 #if WHERETRACE_ENABLED /* 0x8 */ 2170 if( sqlite3WhereTrace & 0x8 ){ 2171 sqlite3DebugPrintf(" skip: "); 2172 whereLoopPrint(pTemplate, pBuilder->pWC); 2173 } 2174 #endif 2175 return SQLITE_OK; 2176 }else{ 2177 p = *ppPrev; 2178 } 2179 2180 /* If we reach this point it means that either p[] should be overwritten 2181 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2182 ** WhereLoop and insert it. 2183 */ 2184 #if WHERETRACE_ENABLED /* 0x8 */ 2185 if( sqlite3WhereTrace & 0x8 ){ 2186 if( p!=0 ){ 2187 sqlite3DebugPrintf("replace: "); 2188 whereLoopPrint(p, pBuilder->pWC); 2189 sqlite3DebugPrintf(" with: "); 2190 }else{ 2191 sqlite3DebugPrintf(" add: "); 2192 } 2193 whereLoopPrint(pTemplate, pBuilder->pWC); 2194 } 2195 #endif 2196 if( p==0 ){ 2197 /* Allocate a new WhereLoop to add to the end of the list */ 2198 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2199 if( p==0 ) return SQLITE_NOMEM_BKPT; 2200 whereLoopInit(p); 2201 p->pNextLoop = 0; 2202 }else{ 2203 /* We will be overwriting WhereLoop p[]. But before we do, first 2204 ** go through the rest of the list and delete any other entries besides 2205 ** p[] that are also supplated by pTemplate */ 2206 WhereLoop **ppTail = &p->pNextLoop; 2207 WhereLoop *pToDel; 2208 while( *ppTail ){ 2209 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2210 if( ppTail==0 ) break; 2211 pToDel = *ppTail; 2212 if( pToDel==0 ) break; 2213 *ppTail = pToDel->pNextLoop; 2214 #if WHERETRACE_ENABLED /* 0x8 */ 2215 if( sqlite3WhereTrace & 0x8 ){ 2216 sqlite3DebugPrintf(" delete: "); 2217 whereLoopPrint(pToDel, pBuilder->pWC); 2218 } 2219 #endif 2220 whereLoopDelete(db, pToDel); 2221 } 2222 } 2223 rc = whereLoopXfer(db, p, pTemplate); 2224 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2225 Index *pIndex = p->u.btree.pIndex; 2226 if( pIndex && pIndex->idxType==SQLITE_IDXTYPE_IPK ){ 2227 p->u.btree.pIndex = 0; 2228 } 2229 } 2230 return rc; 2231 } 2232 2233 /* 2234 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2235 ** WHERE clause that reference the loop but which are not used by an 2236 ** index. 2237 * 2238 ** For every WHERE clause term that is not used by the index 2239 ** and which has a truth probability assigned by one of the likelihood(), 2240 ** likely(), or unlikely() SQL functions, reduce the estimated number 2241 ** of output rows by the probability specified. 2242 ** 2243 ** TUNING: For every WHERE clause term that is not used by the index 2244 ** and which does not have an assigned truth probability, heuristics 2245 ** described below are used to try to estimate the truth probability. 2246 ** TODO --> Perhaps this is something that could be improved by better 2247 ** table statistics. 2248 ** 2249 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2250 ** value corresponds to -1 in LogEst notation, so this means decrement 2251 ** the WhereLoop.nOut field for every such WHERE clause term. 2252 ** 2253 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2254 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2255 ** final output row estimate is no greater than 1/4 of the total number 2256 ** of rows in the table. In other words, assume that x==EXPR will filter 2257 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2258 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2259 ** on the "x" column and so in that case only cap the output row estimate 2260 ** at 1/2 instead of 1/4. 2261 */ 2262 static void whereLoopOutputAdjust( 2263 WhereClause *pWC, /* The WHERE clause */ 2264 WhereLoop *pLoop, /* The loop to adjust downward */ 2265 LogEst nRow /* Number of rows in the entire table */ 2266 ){ 2267 WhereTerm *pTerm, *pX; 2268 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2269 int i, j, k; 2270 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2271 2272 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2273 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ 2274 assert( pTerm!=0 ); 2275 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; 2276 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2277 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2278 for(j=pLoop->nLTerm-1; j>=0; j--){ 2279 pX = pLoop->aLTerm[j]; 2280 if( pX==0 ) continue; 2281 if( pX==pTerm ) break; 2282 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2283 } 2284 if( j<0 ){ 2285 if( pTerm->truthProb<=0 ){ 2286 /* If a truth probability is specified using the likelihood() hints, 2287 ** then use the probability provided by the application. */ 2288 pLoop->nOut += pTerm->truthProb; 2289 }else{ 2290 /* In the absence of explicit truth probabilities, use heuristics to 2291 ** guess a reasonable truth probability. */ 2292 pLoop->nOut--; 2293 if( pTerm->eOperator&(WO_EQ|WO_IS) ){ 2294 Expr *pRight = pTerm->pExpr->pRight; 2295 testcase( pTerm->pExpr->op==TK_IS ); 2296 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2297 k = 10; 2298 }else{ 2299 k = 20; 2300 } 2301 if( iReduce<k ) iReduce = k; 2302 } 2303 } 2304 } 2305 } 2306 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce; 2307 } 2308 2309 /* 2310 ** Term pTerm is a vector range comparison operation. The first comparison 2311 ** in the vector can be optimized using column nEq of the index. This 2312 ** function returns the total number of vector elements that can be used 2313 ** as part of the range comparison. 2314 ** 2315 ** For example, if the query is: 2316 ** 2317 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2318 ** 2319 ** and the index: 2320 ** 2321 ** CREATE INDEX ... ON (a, b, c, d, e) 2322 ** 2323 ** then this function would be invoked with nEq=1. The value returned in 2324 ** this case is 3. 2325 */ 2326 static int whereRangeVectorLen( 2327 Parse *pParse, /* Parsing context */ 2328 int iCur, /* Cursor open on pIdx */ 2329 Index *pIdx, /* The index to be used for a inequality constraint */ 2330 int nEq, /* Number of prior equality constraints on same index */ 2331 WhereTerm *pTerm /* The vector inequality constraint */ 2332 ){ 2333 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2334 int i; 2335 2336 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2337 for(i=1; i<nCmp; i++){ 2338 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2339 ** of the index. If not, exit the loop. */ 2340 char aff; /* Comparison affinity */ 2341 char idxaff = 0; /* Indexed columns affinity */ 2342 CollSeq *pColl; /* Comparison collation sequence */ 2343 Expr *pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2344 Expr *pRhs = pTerm->pExpr->pRight; 2345 if( pRhs->flags & EP_xIsSelect ){ 2346 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2347 }else{ 2348 pRhs = pRhs->x.pList->a[i].pExpr; 2349 } 2350 2351 /* Check that the LHS of the comparison is a column reference to 2352 ** the right column of the right source table. And that the sort 2353 ** order of the index column is the same as the sort order of the 2354 ** leftmost index column. */ 2355 if( pLhs->op!=TK_COLUMN 2356 || pLhs->iTable!=iCur 2357 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2358 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2359 ){ 2360 break; 2361 } 2362 2363 testcase( pLhs->iColumn==XN_ROWID ); 2364 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2365 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2366 if( aff!=idxaff ) break; 2367 2368 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2369 if( pColl==0 ) break; 2370 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2371 } 2372 return i; 2373 } 2374 2375 /* 2376 ** Adjust the cost C by the costMult facter T. This only occurs if 2377 ** compiled with -DSQLITE_ENABLE_COSTMULT 2378 */ 2379 #ifdef SQLITE_ENABLE_COSTMULT 2380 # define ApplyCostMultiplier(C,T) C += T 2381 #else 2382 # define ApplyCostMultiplier(C,T) 2383 #endif 2384 2385 /* 2386 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2387 ** index pIndex. Try to match one more. 2388 ** 2389 ** When this function is called, pBuilder->pNew->nOut contains the 2390 ** number of rows expected to be visited by filtering using the nEq 2391 ** terms only. If it is modified, this value is restored before this 2392 ** function returns. 2393 ** 2394 ** If pProbe->idxType==SQLITE_IDXTYPE_IPK, that means pIndex is 2395 ** a fake index used for the INTEGER PRIMARY KEY. 2396 */ 2397 static int whereLoopAddBtreeIndex( 2398 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2399 struct SrcList_item *pSrc, /* FROM clause term being analyzed */ 2400 Index *pProbe, /* An index on pSrc */ 2401 LogEst nInMul /* log(Number of iterations due to IN) */ 2402 ){ 2403 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2404 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2405 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2406 WhereLoop *pNew; /* Template WhereLoop under construction */ 2407 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2408 int opMask; /* Valid operators for constraints */ 2409 WhereScan scan; /* Iterator for WHERE terms */ 2410 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2411 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2412 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2413 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2414 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2415 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2416 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2417 LogEst saved_nOut; /* Original value of pNew->nOut */ 2418 int rc = SQLITE_OK; /* Return code */ 2419 LogEst rSize; /* Number of rows in the table */ 2420 LogEst rLogSize; /* Logarithm of table size */ 2421 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2422 2423 pNew = pBuilder->pNew; 2424 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2425 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d\n", 2426 pProbe->pTable->zName,pProbe->zName, pNew->u.btree.nEq)); 2427 2428 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2429 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2430 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2431 opMask = WO_LT|WO_LE; 2432 }else{ 2433 assert( pNew->u.btree.nBtm==0 ); 2434 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2435 } 2436 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2437 2438 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2439 2440 saved_nEq = pNew->u.btree.nEq; 2441 saved_nBtm = pNew->u.btree.nBtm; 2442 saved_nTop = pNew->u.btree.nTop; 2443 saved_nSkip = pNew->nSkip; 2444 saved_nLTerm = pNew->nLTerm; 2445 saved_wsFlags = pNew->wsFlags; 2446 saved_prereq = pNew->prereq; 2447 saved_nOut = pNew->nOut; 2448 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2449 opMask, pProbe); 2450 pNew->rSetup = 0; 2451 rSize = pProbe->aiRowLogEst[0]; 2452 rLogSize = estLog(rSize); 2453 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2454 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2455 LogEst rCostIdx; 2456 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2457 int nIn = 0; 2458 #ifdef SQLITE_ENABLE_STAT4 2459 int nRecValid = pBuilder->nRecValid; 2460 #endif 2461 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2462 && indexColumnNotNull(pProbe, saved_nEq) 2463 ){ 2464 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2465 } 2466 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2467 2468 /* Do not allow the upper bound of a LIKE optimization range constraint 2469 ** to mix with a lower range bound from some other source */ 2470 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2471 2472 /* Do not allow constraints from the WHERE clause to be used by the 2473 ** right table of a LEFT JOIN. Only constraints in the ON clause are 2474 ** allowed */ 2475 if( (pSrc->fg.jointype & JT_LEFT)!=0 2476 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2477 ){ 2478 continue; 2479 } 2480 2481 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2482 pBuilder->bldFlags |= SQLITE_BLDF_UNIQUE; 2483 }else{ 2484 pBuilder->bldFlags |= SQLITE_BLDF_INDEXED; 2485 } 2486 pNew->wsFlags = saved_wsFlags; 2487 pNew->u.btree.nEq = saved_nEq; 2488 pNew->u.btree.nBtm = saved_nBtm; 2489 pNew->u.btree.nTop = saved_nTop; 2490 pNew->nLTerm = saved_nLTerm; 2491 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2492 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2493 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2494 2495 assert( nInMul==0 2496 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2497 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2498 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2499 ); 2500 2501 if( eOp & WO_IN ){ 2502 Expr *pExpr = pTerm->pExpr; 2503 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 2504 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2505 int i; 2506 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2507 2508 /* The expression may actually be of the form (x, y) IN (SELECT...). 2509 ** In this case there is a separate term for each of (x) and (y). 2510 ** However, the nIn multiplier should only be applied once, not once 2511 ** for each such term. The following loop checks that pTerm is the 2512 ** first such term in use, and sets nIn back to 0 if it is not. */ 2513 for(i=0; i<pNew->nLTerm-1; i++){ 2514 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2515 } 2516 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2517 /* "x IN (value, value, ...)" */ 2518 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2519 assert( nIn>0 ); /* RHS always has 2 or more terms... The parser 2520 ** changes "x IN (?)" into "x=?". */ 2521 } 2522 if( pProbe->hasStat1 ){ 2523 LogEst M, logK, safetyMargin; 2524 /* Let: 2525 ** N = the total number of rows in the table 2526 ** K = the number of entries on the RHS of the IN operator 2527 ** M = the number of rows in the table that match terms to the 2528 ** to the left in the same index. If the IN operator is on 2529 ** the left-most index column, M==N. 2530 ** 2531 ** Given the definitions above, it is better to omit the IN operator 2532 ** from the index lookup and instead do a scan of the M elements, 2533 ** testing each scanned row against the IN operator separately, if: 2534 ** 2535 ** M*log(K) < K*log(N) 2536 ** 2537 ** Our estimates for M, K, and N might be inaccurate, so we build in 2538 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator 2539 ** with the index, as using an index has better worst-case behavior. 2540 ** If we do not have real sqlite_stat1 data, always prefer to use 2541 ** the index. 2542 */ 2543 M = pProbe->aiRowLogEst[saved_nEq]; 2544 logK = estLog(nIn); 2545 safetyMargin = 10; /* TUNING: extra weight for indexed IN */ 2546 if( M + logK + safetyMargin < nIn + rLogSize ){ 2547 WHERETRACE(0x40, 2548 ("Scan preferred over IN operator on column %d of \"%s\" (%d<%d)\n", 2549 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2550 continue; 2551 }else{ 2552 WHERETRACE(0x40, 2553 ("IN operator preferred on column %d of \"%s\" (%d>=%d)\n", 2554 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2555 } 2556 } 2557 pNew->wsFlags |= WHERE_COLUMN_IN; 2558 }else if( eOp & (WO_EQ|WO_IS) ){ 2559 int iCol = pProbe->aiColumn[saved_nEq]; 2560 pNew->wsFlags |= WHERE_COLUMN_EQ; 2561 assert( saved_nEq==pNew->u.btree.nEq ); 2562 if( iCol==XN_ROWID 2563 || (iCol>=0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2564 ){ 2565 if( iCol==XN_ROWID || pProbe->uniqNotNull 2566 || (pProbe->nKeyCol==1 && pProbe->onError && eOp==WO_EQ) 2567 ){ 2568 pNew->wsFlags |= WHERE_ONEROW; 2569 }else{ 2570 pNew->wsFlags |= WHERE_UNQ_WANTED; 2571 } 2572 } 2573 }else if( eOp & WO_ISNULL ){ 2574 pNew->wsFlags |= WHERE_COLUMN_NULL; 2575 }else if( eOp & (WO_GT|WO_GE) ){ 2576 testcase( eOp & WO_GT ); 2577 testcase( eOp & WO_GE ); 2578 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2579 pNew->u.btree.nBtm = whereRangeVectorLen( 2580 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2581 ); 2582 pBtm = pTerm; 2583 pTop = 0; 2584 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2585 /* Range contraints that come from the LIKE optimization are 2586 ** always used in pairs. */ 2587 pTop = &pTerm[1]; 2588 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2589 assert( pTop->wtFlags & TERM_LIKEOPT ); 2590 assert( pTop->eOperator==WO_LT ); 2591 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2592 pNew->aLTerm[pNew->nLTerm++] = pTop; 2593 pNew->wsFlags |= WHERE_TOP_LIMIT; 2594 pNew->u.btree.nTop = 1; 2595 } 2596 }else{ 2597 assert( eOp & (WO_LT|WO_LE) ); 2598 testcase( eOp & WO_LT ); 2599 testcase( eOp & WO_LE ); 2600 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2601 pNew->u.btree.nTop = whereRangeVectorLen( 2602 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2603 ); 2604 pTop = pTerm; 2605 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2606 pNew->aLTerm[pNew->nLTerm-2] : 0; 2607 } 2608 2609 /* At this point pNew->nOut is set to the number of rows expected to 2610 ** be visited by the index scan before considering term pTerm, or the 2611 ** values of nIn and nInMul. In other words, assuming that all 2612 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2613 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2614 assert( pNew->nOut==saved_nOut ); 2615 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2616 /* Adjust nOut using stat4 data. Or, if there is no stat4 2617 ** data, using some other estimate. */ 2618 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2619 }else{ 2620 int nEq = ++pNew->u.btree.nEq; 2621 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2622 2623 assert( pNew->nOut==saved_nOut ); 2624 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2625 assert( (eOp & WO_IN) || nIn==0 ); 2626 testcase( eOp & WO_IN ); 2627 pNew->nOut += pTerm->truthProb; 2628 pNew->nOut -= nIn; 2629 }else{ 2630 #ifdef SQLITE_ENABLE_STAT4 2631 tRowcnt nOut = 0; 2632 if( nInMul==0 2633 && pProbe->nSample 2634 && pNew->u.btree.nEq<=pProbe->nSampleCol 2635 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) 2636 && OptimizationEnabled(db, SQLITE_Stat4) 2637 ){ 2638 Expr *pExpr = pTerm->pExpr; 2639 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2640 testcase( eOp & WO_EQ ); 2641 testcase( eOp & WO_IS ); 2642 testcase( eOp & WO_ISNULL ); 2643 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2644 }else{ 2645 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2646 } 2647 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2648 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2649 if( nOut ){ 2650 pNew->nOut = sqlite3LogEst(nOut); 2651 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2652 pNew->nOut -= nIn; 2653 } 2654 } 2655 if( nOut==0 ) 2656 #endif 2657 { 2658 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2659 if( eOp & WO_ISNULL ){ 2660 /* TUNING: If there is no likelihood() value, assume that a 2661 ** "col IS NULL" expression matches twice as many rows 2662 ** as (col=?). */ 2663 pNew->nOut += 10; 2664 } 2665 } 2666 } 2667 } 2668 2669 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2670 ** it to pNew->rRun, which is currently set to the cost of the index 2671 ** seek only. Then, if this is a non-covering index, add the cost of 2672 ** visiting the rows in the main table. */ 2673 assert( pSrc->pTab->szTabRow>0 ); 2674 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2675 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2676 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2677 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2678 } 2679 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2680 2681 nOutUnadjusted = pNew->nOut; 2682 pNew->rRun += nInMul + nIn; 2683 pNew->nOut += nInMul + nIn; 2684 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 2685 rc = whereLoopInsert(pBuilder, pNew); 2686 2687 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2688 pNew->nOut = saved_nOut; 2689 }else{ 2690 pNew->nOut = nOutUnadjusted; 2691 } 2692 2693 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 2694 && pNew->u.btree.nEq<pProbe->nColumn 2695 ){ 2696 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 2697 } 2698 pNew->nOut = saved_nOut; 2699 #ifdef SQLITE_ENABLE_STAT4 2700 pBuilder->nRecValid = nRecValid; 2701 #endif 2702 } 2703 pNew->prereq = saved_prereq; 2704 pNew->u.btree.nEq = saved_nEq; 2705 pNew->u.btree.nBtm = saved_nBtm; 2706 pNew->u.btree.nTop = saved_nTop; 2707 pNew->nSkip = saved_nSkip; 2708 pNew->wsFlags = saved_wsFlags; 2709 pNew->nOut = saved_nOut; 2710 pNew->nLTerm = saved_nLTerm; 2711 2712 /* Consider using a skip-scan if there are no WHERE clause constraints 2713 ** available for the left-most terms of the index, and if the average 2714 ** number of repeats in the left-most terms is at least 18. 2715 ** 2716 ** The magic number 18 is selected on the basis that scanning 17 rows 2717 ** is almost always quicker than an index seek (even though if the index 2718 ** contains fewer than 2^17 rows we assume otherwise in other parts of 2719 ** the code). And, even if it is not, it should not be too much slower. 2720 ** On the other hand, the extra seeks could end up being significantly 2721 ** more expensive. */ 2722 assert( 42==sqlite3LogEst(18) ); 2723 if( saved_nEq==saved_nSkip 2724 && saved_nEq+1<pProbe->nKeyCol 2725 && pProbe->noSkipScan==0 2726 && OptimizationEnabled(db, SQLITE_SkipScan) 2727 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 2728 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 2729 ){ 2730 LogEst nIter; 2731 pNew->u.btree.nEq++; 2732 pNew->nSkip++; 2733 pNew->aLTerm[pNew->nLTerm++] = 0; 2734 pNew->wsFlags |= WHERE_SKIPSCAN; 2735 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 2736 pNew->nOut -= nIter; 2737 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 2738 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 2739 nIter += 5; 2740 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 2741 pNew->nOut = saved_nOut; 2742 pNew->u.btree.nEq = saved_nEq; 2743 pNew->nSkip = saved_nSkip; 2744 pNew->wsFlags = saved_wsFlags; 2745 } 2746 2747 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n", 2748 pProbe->pTable->zName, pProbe->zName, saved_nEq, rc)); 2749 return rc; 2750 } 2751 2752 /* 2753 ** Return True if it is possible that pIndex might be useful in 2754 ** implementing the ORDER BY clause in pBuilder. 2755 ** 2756 ** Return False if pBuilder does not contain an ORDER BY clause or 2757 ** if there is no way for pIndex to be useful in implementing that 2758 ** ORDER BY clause. 2759 */ 2760 static int indexMightHelpWithOrderBy( 2761 WhereLoopBuilder *pBuilder, 2762 Index *pIndex, 2763 int iCursor 2764 ){ 2765 ExprList *pOB; 2766 ExprList *aColExpr; 2767 int ii, jj; 2768 2769 if( pIndex->bUnordered ) return 0; 2770 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 2771 for(ii=0; ii<pOB->nExpr; ii++){ 2772 Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr); 2773 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 2774 if( pExpr->iColumn<0 ) return 1; 2775 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2776 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 2777 } 2778 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 2779 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2780 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 2781 if( sqlite3ExprCompareSkip(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 2782 return 1; 2783 } 2784 } 2785 } 2786 } 2787 return 0; 2788 } 2789 2790 /* Check to see if a partial index with pPartIndexWhere can be used 2791 ** in the current query. Return true if it can be and false if not. 2792 */ 2793 static int whereUsablePartialIndex(int iTab, WhereClause *pWC, Expr *pWhere){ 2794 int i; 2795 WhereTerm *pTerm; 2796 Parse *pParse = pWC->pWInfo->pParse; 2797 while( pWhere->op==TK_AND ){ 2798 if( !whereUsablePartialIndex(iTab,pWC,pWhere->pLeft) ) return 0; 2799 pWhere = pWhere->pRight; 2800 } 2801 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 2802 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2803 Expr *pExpr = pTerm->pExpr; 2804 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 2805 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 2806 ){ 2807 return 1; 2808 } 2809 } 2810 return 0; 2811 } 2812 2813 /* 2814 ** Add all WhereLoop objects for a single table of the join where the table 2815 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 2816 ** a b-tree table, not a virtual table. 2817 ** 2818 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 2819 ** are calculated as follows: 2820 ** 2821 ** For a full scan, assuming the table (or index) contains nRow rows: 2822 ** 2823 ** cost = nRow * 3.0 // full-table scan 2824 ** cost = nRow * K // scan of covering index 2825 ** cost = nRow * (K+3.0) // scan of non-covering index 2826 ** 2827 ** where K is a value between 1.1 and 3.0 set based on the relative 2828 ** estimated average size of the index and table records. 2829 ** 2830 ** For an index scan, where nVisit is the number of index rows visited 2831 ** by the scan, and nSeek is the number of seek operations required on 2832 ** the index b-tree: 2833 ** 2834 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 2835 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 2836 ** 2837 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 2838 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 2839 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 2840 ** 2841 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 2842 ** of uncertainty. For this reason, scoring is designed to pick plans that 2843 ** "do the least harm" if the estimates are inaccurate. For example, a 2844 ** log(nRow) factor is omitted from a non-covering index scan in order to 2845 ** bias the scoring in favor of using an index, since the worst-case 2846 ** performance of using an index is far better than the worst-case performance 2847 ** of a full table scan. 2848 */ 2849 static int whereLoopAddBtree( 2850 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2851 Bitmask mPrereq /* Extra prerequesites for using this table */ 2852 ){ 2853 WhereInfo *pWInfo; /* WHERE analysis context */ 2854 Index *pProbe; /* An index we are evaluating */ 2855 Index sPk; /* A fake index object for the primary key */ 2856 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 2857 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 2858 SrcList *pTabList; /* The FROM clause */ 2859 struct SrcList_item *pSrc; /* The FROM clause btree term to add */ 2860 WhereLoop *pNew; /* Template WhereLoop object */ 2861 int rc = SQLITE_OK; /* Return code */ 2862 int iSortIdx = 1; /* Index number */ 2863 int b; /* A boolean value */ 2864 LogEst rSize; /* number of rows in the table */ 2865 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 2866 WhereClause *pWC; /* The parsed WHERE clause */ 2867 Table *pTab; /* Table being queried */ 2868 2869 pNew = pBuilder->pNew; 2870 pWInfo = pBuilder->pWInfo; 2871 pTabList = pWInfo->pTabList; 2872 pSrc = pTabList->a + pNew->iTab; 2873 pTab = pSrc->pTab; 2874 pWC = pBuilder->pWC; 2875 assert( !IsVirtual(pSrc->pTab) ); 2876 2877 if( pSrc->pIBIndex ){ 2878 /* An INDEXED BY clause specifies a particular index to use */ 2879 pProbe = pSrc->pIBIndex; 2880 }else if( !HasRowid(pTab) ){ 2881 pProbe = pTab->pIndex; 2882 }else{ 2883 /* There is no INDEXED BY clause. Create a fake Index object in local 2884 ** variable sPk to represent the rowid primary key index. Make this 2885 ** fake index the first in a chain of Index objects with all of the real 2886 ** indices to follow */ 2887 Index *pFirst; /* First of real indices on the table */ 2888 memset(&sPk, 0, sizeof(Index)); 2889 sPk.nKeyCol = 1; 2890 sPk.nColumn = 1; 2891 sPk.aiColumn = &aiColumnPk; 2892 sPk.aiRowLogEst = aiRowEstPk; 2893 sPk.onError = OE_Replace; 2894 sPk.pTable = pTab; 2895 sPk.szIdxRow = pTab->szTabRow; 2896 sPk.idxType = SQLITE_IDXTYPE_IPK; 2897 aiRowEstPk[0] = pTab->nRowLogEst; 2898 aiRowEstPk[1] = 0; 2899 pFirst = pSrc->pTab->pIndex; 2900 if( pSrc->fg.notIndexed==0 ){ 2901 /* The real indices of the table are only considered if the 2902 ** NOT INDEXED qualifier is omitted from the FROM clause */ 2903 sPk.pNext = pFirst; 2904 } 2905 pProbe = &sPk; 2906 } 2907 rSize = pTab->nRowLogEst; 2908 rLogSize = estLog(rSize); 2909 2910 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 2911 /* Automatic indexes */ 2912 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 2913 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 2914 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 2915 && pSrc->pIBIndex==0 /* Has no INDEXED BY clause */ 2916 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 2917 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 2918 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 2919 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 2920 ){ 2921 /* Generate auto-index WhereLoops */ 2922 WhereTerm *pTerm; 2923 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 2924 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 2925 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2926 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 2927 pNew->u.btree.nEq = 1; 2928 pNew->nSkip = 0; 2929 pNew->u.btree.pIndex = 0; 2930 pNew->nLTerm = 1; 2931 pNew->aLTerm[0] = pTerm; 2932 /* TUNING: One-time cost for computing the automatic index is 2933 ** estimated to be X*N*log2(N) where N is the number of rows in 2934 ** the table being indexed and where X is 7 (LogEst=28) for normal 2935 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value 2936 ** of X is smaller for views and subqueries so that the query planner 2937 ** will be more aggressive about generating automatic indexes for 2938 ** those objects, since there is no opportunity to add schema 2939 ** indexes on subqueries and views. */ 2940 pNew->rSetup = rLogSize + rSize; 2941 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ 2942 pNew->rSetup += 28; 2943 }else{ 2944 pNew->rSetup -= 10; 2945 } 2946 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 2947 if( pNew->rSetup<0 ) pNew->rSetup = 0; 2948 /* TUNING: Each index lookup yields 20 rows in the table. This 2949 ** is more than the usual guess of 10 rows, since we have no way 2950 ** of knowing how selective the index will ultimately be. It would 2951 ** not be unreasonable to make this value much larger. */ 2952 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 2953 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 2954 pNew->wsFlags = WHERE_AUTO_INDEX; 2955 pNew->prereq = mPrereq | pTerm->prereqRight; 2956 rc = whereLoopInsert(pBuilder, pNew); 2957 } 2958 } 2959 } 2960 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 2961 2962 /* Loop over all indices. If there was an INDEXED BY clause, then only 2963 ** consider index pProbe. */ 2964 for(; rc==SQLITE_OK && pProbe; 2965 pProbe=(pSrc->pIBIndex ? 0 : pProbe->pNext), iSortIdx++ 2966 ){ 2967 if( pProbe->pPartIdxWhere!=0 2968 && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){ 2969 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 2970 continue; /* Partial index inappropriate for this query */ 2971 } 2972 if( pProbe->bNoQuery ) continue; 2973 rSize = pProbe->aiRowLogEst[0]; 2974 pNew->u.btree.nEq = 0; 2975 pNew->u.btree.nBtm = 0; 2976 pNew->u.btree.nTop = 0; 2977 pNew->nSkip = 0; 2978 pNew->nLTerm = 0; 2979 pNew->iSortIdx = 0; 2980 pNew->rSetup = 0; 2981 pNew->prereq = mPrereq; 2982 pNew->nOut = rSize; 2983 pNew->u.btree.pIndex = pProbe; 2984 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 2985 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 2986 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 2987 if( pProbe->idxType==SQLITE_IDXTYPE_IPK ){ 2988 /* Integer primary key index */ 2989 pNew->wsFlags = WHERE_IPK; 2990 2991 /* Full table scan */ 2992 pNew->iSortIdx = b ? iSortIdx : 0; 2993 /* TUNING: Cost of full table scan is (N*3.0). */ 2994 pNew->rRun = rSize + 16; 2995 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2996 whereLoopOutputAdjust(pWC, pNew, rSize); 2997 rc = whereLoopInsert(pBuilder, pNew); 2998 pNew->nOut = rSize; 2999 if( rc ) break; 3000 }else{ 3001 Bitmask m; 3002 if( pProbe->isCovering ){ 3003 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 3004 m = 0; 3005 }else{ 3006 m = pSrc->colUsed & pProbe->colNotIdxed; 3007 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 3008 } 3009 3010 /* Full scan via index */ 3011 if( b 3012 || !HasRowid(pTab) 3013 || pProbe->pPartIdxWhere!=0 3014 || ( m==0 3015 && pProbe->bUnordered==0 3016 && (pProbe->szIdxRow<pTab->szTabRow) 3017 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 3018 && sqlite3GlobalConfig.bUseCis 3019 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 3020 ) 3021 ){ 3022 pNew->iSortIdx = b ? iSortIdx : 0; 3023 3024 /* The cost of visiting the index rows is N*K, where K is 3025 ** between 1.1 and 3.0, depending on the relative sizes of the 3026 ** index and table rows. */ 3027 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 3028 if( m!=0 ){ 3029 /* If this is a non-covering index scan, add in the cost of 3030 ** doing table lookups. The cost will be 3x the number of 3031 ** lookups. Take into account WHERE clause terms that can be 3032 ** satisfied using just the index, and that do not require a 3033 ** table lookup. */ 3034 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 3035 int ii; 3036 int iCur = pSrc->iCursor; 3037 WhereClause *pWC2 = &pWInfo->sWC; 3038 for(ii=0; ii<pWC2->nTerm; ii++){ 3039 WhereTerm *pTerm = &pWC2->a[ii]; 3040 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 3041 break; 3042 } 3043 /* pTerm can be evaluated using just the index. So reduce 3044 ** the expected number of table lookups accordingly */ 3045 if( pTerm->truthProb<=0 ){ 3046 nLookup += pTerm->truthProb; 3047 }else{ 3048 nLookup--; 3049 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 3050 } 3051 } 3052 3053 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 3054 } 3055 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3056 whereLoopOutputAdjust(pWC, pNew, rSize); 3057 rc = whereLoopInsert(pBuilder, pNew); 3058 pNew->nOut = rSize; 3059 if( rc ) break; 3060 } 3061 } 3062 3063 pBuilder->bldFlags = 0; 3064 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 3065 if( pBuilder->bldFlags==SQLITE_BLDF_INDEXED ){ 3066 /* If a non-unique index is used, or if a prefix of the key for 3067 ** unique index is used (making the index functionally non-unique) 3068 ** then the sqlite_stat1 data becomes important for scoring the 3069 ** plan */ 3070 pTab->tabFlags |= TF_StatsUsed; 3071 } 3072 #ifdef SQLITE_ENABLE_STAT4 3073 sqlite3Stat4ProbeFree(pBuilder->pRec); 3074 pBuilder->nRecValid = 0; 3075 pBuilder->pRec = 0; 3076 #endif 3077 } 3078 return rc; 3079 } 3080 3081 #ifndef SQLITE_OMIT_VIRTUALTABLE 3082 3083 /* 3084 ** Argument pIdxInfo is already populated with all constraints that may 3085 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 3086 ** function marks a subset of those constraints usable, invokes the 3087 ** xBestIndex method and adds the returned plan to pBuilder. 3088 ** 3089 ** A constraint is marked usable if: 3090 ** 3091 ** * Argument mUsable indicates that its prerequisites are available, and 3092 ** 3093 ** * It is not one of the operators specified in the mExclude mask passed 3094 ** as the fourth argument (which in practice is either WO_IN or 0). 3095 ** 3096 ** Argument mPrereq is a mask of tables that must be scanned before the 3097 ** virtual table in question. These are added to the plans prerequisites 3098 ** before it is added to pBuilder. 3099 ** 3100 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3101 ** uses one or more WO_IN terms, or false otherwise. 3102 */ 3103 static int whereLoopAddVirtualOne( 3104 WhereLoopBuilder *pBuilder, 3105 Bitmask mPrereq, /* Mask of tables that must be used. */ 3106 Bitmask mUsable, /* Mask of usable tables */ 3107 u16 mExclude, /* Exclude terms using these operators */ 3108 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3109 u16 mNoOmit, /* Do not omit these constraints */ 3110 int *pbIn /* OUT: True if plan uses an IN(...) op */ 3111 ){ 3112 WhereClause *pWC = pBuilder->pWC; 3113 struct sqlite3_index_constraint *pIdxCons; 3114 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3115 int i; 3116 int mxTerm; 3117 int rc = SQLITE_OK; 3118 WhereLoop *pNew = pBuilder->pNew; 3119 Parse *pParse = pBuilder->pWInfo->pParse; 3120 struct SrcList_item *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3121 int nConstraint = pIdxInfo->nConstraint; 3122 3123 assert( (mUsable & mPrereq)==mPrereq ); 3124 *pbIn = 0; 3125 pNew->prereq = mPrereq; 3126 3127 /* Set the usable flag on the subset of constraints identified by 3128 ** arguments mUsable and mExclude. */ 3129 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3130 for(i=0; i<nConstraint; i++, pIdxCons++){ 3131 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3132 pIdxCons->usable = 0; 3133 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3134 && (pTerm->eOperator & mExclude)==0 3135 ){ 3136 pIdxCons->usable = 1; 3137 } 3138 } 3139 3140 /* Initialize the output fields of the sqlite3_index_info structure */ 3141 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3142 assert( pIdxInfo->needToFreeIdxStr==0 ); 3143 pIdxInfo->idxStr = 0; 3144 pIdxInfo->idxNum = 0; 3145 pIdxInfo->orderByConsumed = 0; 3146 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3147 pIdxInfo->estimatedRows = 25; 3148 pIdxInfo->idxFlags = 0; 3149 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3150 3151 /* Invoke the virtual table xBestIndex() method */ 3152 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3153 if( rc ){ 3154 if( rc==SQLITE_CONSTRAINT ){ 3155 /* If the xBestIndex method returns SQLITE_CONSTRAINT, that means 3156 ** that the particular combination of parameters provided is unusable. 3157 ** Make no entries in the loop table. 3158 */ 3159 WHERETRACE(0xffff, (" ^^^^--- non-viable plan rejected!\n")); 3160 return SQLITE_OK; 3161 } 3162 return rc; 3163 } 3164 3165 mxTerm = -1; 3166 assert( pNew->nLSlot>=nConstraint ); 3167 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 3168 pNew->u.vtab.omitMask = 0; 3169 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3170 for(i=0; i<nConstraint; i++, pIdxCons++){ 3171 int iTerm; 3172 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3173 WhereTerm *pTerm; 3174 int j = pIdxCons->iTermOffset; 3175 if( iTerm>=nConstraint 3176 || j<0 3177 || j>=pWC->nTerm 3178 || pNew->aLTerm[iTerm]!=0 3179 || pIdxCons->usable==0 3180 ){ 3181 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3182 testcase( pIdxInfo->needToFreeIdxStr ); 3183 return SQLITE_ERROR; 3184 } 3185 testcase( iTerm==nConstraint-1 ); 3186 testcase( j==0 ); 3187 testcase( j==pWC->nTerm-1 ); 3188 pTerm = &pWC->a[j]; 3189 pNew->prereq |= pTerm->prereqRight; 3190 assert( iTerm<pNew->nLSlot ); 3191 pNew->aLTerm[iTerm] = pTerm; 3192 if( iTerm>mxTerm ) mxTerm = iTerm; 3193 testcase( iTerm==15 ); 3194 testcase( iTerm==16 ); 3195 if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<<iTerm; 3196 if( (pTerm->eOperator & WO_IN)!=0 ){ 3197 /* A virtual table that is constrained by an IN clause may not 3198 ** consume the ORDER BY clause because (1) the order of IN terms 3199 ** is not necessarily related to the order of output terms and 3200 ** (2) Multiple outputs from a single IN value will not merge 3201 ** together. */ 3202 pIdxInfo->orderByConsumed = 0; 3203 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3204 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3205 } 3206 } 3207 } 3208 pNew->u.vtab.omitMask &= ~mNoOmit; 3209 3210 pNew->nLTerm = mxTerm+1; 3211 for(i=0; i<=mxTerm; i++){ 3212 if( pNew->aLTerm[i]==0 ){ 3213 /* The non-zero argvIdx values must be contiguous. Raise an 3214 ** error if they are not */ 3215 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3216 testcase( pIdxInfo->needToFreeIdxStr ); 3217 return SQLITE_ERROR; 3218 } 3219 } 3220 assert( pNew->nLTerm<=pNew->nLSlot ); 3221 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3222 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3223 pIdxInfo->needToFreeIdxStr = 0; 3224 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3225 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3226 pIdxInfo->nOrderBy : 0); 3227 pNew->rSetup = 0; 3228 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3229 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3230 3231 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3232 ** that the scan will visit at most one row. Clear it otherwise. */ 3233 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3234 pNew->wsFlags |= WHERE_ONEROW; 3235 }else{ 3236 pNew->wsFlags &= ~WHERE_ONEROW; 3237 } 3238 rc = whereLoopInsert(pBuilder, pNew); 3239 if( pNew->u.vtab.needFree ){ 3240 sqlite3_free(pNew->u.vtab.idxStr); 3241 pNew->u.vtab.needFree = 0; 3242 } 3243 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3244 *pbIn, (sqlite3_uint64)mPrereq, 3245 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3246 3247 return rc; 3248 } 3249 3250 /* 3251 ** If this function is invoked from within an xBestIndex() callback, it 3252 ** returns a pointer to a buffer containing the name of the collation 3253 ** sequence associated with element iCons of the sqlite3_index_info.aConstraint 3254 ** array. Or, if iCons is out of range or there is no active xBestIndex 3255 ** call, return NULL. 3256 */ 3257 const char *sqlite3_vtab_collation(sqlite3_index_info *pIdxInfo, int iCons){ 3258 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3259 const char *zRet = 0; 3260 if( iCons>=0 && iCons<pIdxInfo->nConstraint ){ 3261 CollSeq *pC = 0; 3262 int iTerm = pIdxInfo->aConstraint[iCons].iTermOffset; 3263 Expr *pX = pHidden->pWC->a[iTerm].pExpr; 3264 if( pX->pLeft ){ 3265 pC = sqlite3BinaryCompareCollSeq(pHidden->pParse, pX->pLeft, pX->pRight); 3266 } 3267 zRet = (pC ? pC->zName : sqlite3StrBINARY); 3268 } 3269 return zRet; 3270 } 3271 3272 /* 3273 ** Add all WhereLoop objects for a table of the join identified by 3274 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3275 ** 3276 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3277 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3278 ** entries that occur before the virtual table in the FROM clause and are 3279 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3280 ** mUnusable mask contains all FROM clause entries that occur after the 3281 ** virtual table and are separated from it by at least one LEFT or 3282 ** CROSS JOIN. 3283 ** 3284 ** For example, if the query were: 3285 ** 3286 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3287 ** 3288 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3289 ** 3290 ** All the tables in mPrereq must be scanned before the current virtual 3291 ** table. So any terms for which all prerequisites are satisfied by 3292 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3293 ** Conversely, all tables in mUnusable must be scanned after the current 3294 ** virtual table, so any terms for which the prerequisites overlap with 3295 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3296 */ 3297 static int whereLoopAddVirtual( 3298 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3299 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3300 Bitmask mUnusable /* Tables that must be scanned after this one */ 3301 ){ 3302 int rc = SQLITE_OK; /* Return code */ 3303 WhereInfo *pWInfo; /* WHERE analysis context */ 3304 Parse *pParse; /* The parsing context */ 3305 WhereClause *pWC; /* The WHERE clause */ 3306 struct SrcList_item *pSrc; /* The FROM clause term to search */ 3307 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3308 int nConstraint; /* Number of constraints in p */ 3309 int bIn; /* True if plan uses IN(...) operator */ 3310 WhereLoop *pNew; 3311 Bitmask mBest; /* Tables used by best possible plan */ 3312 u16 mNoOmit; 3313 3314 assert( (mPrereq & mUnusable)==0 ); 3315 pWInfo = pBuilder->pWInfo; 3316 pParse = pWInfo->pParse; 3317 pWC = pBuilder->pWC; 3318 pNew = pBuilder->pNew; 3319 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3320 assert( IsVirtual(pSrc->pTab) ); 3321 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy, 3322 &mNoOmit); 3323 if( p==0 ) return SQLITE_NOMEM_BKPT; 3324 pNew->rSetup = 0; 3325 pNew->wsFlags = WHERE_VIRTUALTABLE; 3326 pNew->nLTerm = 0; 3327 pNew->u.vtab.needFree = 0; 3328 nConstraint = p->nConstraint; 3329 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3330 sqlite3DbFree(pParse->db, p); 3331 return SQLITE_NOMEM_BKPT; 3332 } 3333 3334 /* First call xBestIndex() with all constraints usable. */ 3335 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc->pTab->zName)); 3336 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3337 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn); 3338 3339 /* If the call to xBestIndex() with all terms enabled produced a plan 3340 ** that does not require any source tables (IOW: a plan with mBest==0) 3341 ** and does not use an IN(...) operator, then there is no point in making 3342 ** any further calls to xBestIndex() since they will all return the same 3343 ** result (if the xBestIndex() implementation is sane). */ 3344 if( rc==SQLITE_OK && ((mBest = (pNew->prereq & ~mPrereq))!=0 || bIn) ){ 3345 int seenZero = 0; /* True if a plan with no prereqs seen */ 3346 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3347 Bitmask mPrev = 0; 3348 Bitmask mBestNoIn = 0; 3349 3350 /* If the plan produced by the earlier call uses an IN(...) term, call 3351 ** xBestIndex again, this time with IN(...) terms disabled. */ 3352 if( bIn ){ 3353 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3354 rc = whereLoopAddVirtualOne( 3355 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn); 3356 assert( bIn==0 ); 3357 mBestNoIn = pNew->prereq & ~mPrereq; 3358 if( mBestNoIn==0 ){ 3359 seenZero = 1; 3360 seenZeroNoIN = 1; 3361 } 3362 } 3363 3364 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3365 ** in the set of terms that apply to the current virtual table. */ 3366 while( rc==SQLITE_OK ){ 3367 int i; 3368 Bitmask mNext = ALLBITS; 3369 assert( mNext>0 ); 3370 for(i=0; i<nConstraint; i++){ 3371 Bitmask mThis = ( 3372 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3373 ); 3374 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3375 } 3376 mPrev = mNext; 3377 if( mNext==ALLBITS ) break; 3378 if( mNext==mBest || mNext==mBestNoIn ) continue; 3379 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3380 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3381 rc = whereLoopAddVirtualOne( 3382 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn); 3383 if( pNew->prereq==mPrereq ){ 3384 seenZero = 1; 3385 if( bIn==0 ) seenZeroNoIN = 1; 3386 } 3387 } 3388 3389 /* If the calls to xBestIndex() in the above loop did not find a plan 3390 ** that requires no source tables at all (i.e. one guaranteed to be 3391 ** usable), make a call here with all source tables disabled */ 3392 if( rc==SQLITE_OK && seenZero==0 ){ 3393 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3394 rc = whereLoopAddVirtualOne( 3395 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn); 3396 if( bIn==0 ) seenZeroNoIN = 1; 3397 } 3398 3399 /* If the calls to xBestIndex() have so far failed to find a plan 3400 ** that requires no source tables at all and does not use an IN(...) 3401 ** operator, make a final call to obtain one here. */ 3402 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3403 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3404 rc = whereLoopAddVirtualOne( 3405 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn); 3406 } 3407 } 3408 3409 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3410 sqlite3DbFreeNN(pParse->db, p); 3411 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc->pTab->zName, rc)); 3412 return rc; 3413 } 3414 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3415 3416 /* 3417 ** Add WhereLoop entries to handle OR terms. This works for either 3418 ** btrees or virtual tables. 3419 */ 3420 static int whereLoopAddOr( 3421 WhereLoopBuilder *pBuilder, 3422 Bitmask mPrereq, 3423 Bitmask mUnusable 3424 ){ 3425 WhereInfo *pWInfo = pBuilder->pWInfo; 3426 WhereClause *pWC; 3427 WhereLoop *pNew; 3428 WhereTerm *pTerm, *pWCEnd; 3429 int rc = SQLITE_OK; 3430 int iCur; 3431 WhereClause tempWC; 3432 WhereLoopBuilder sSubBuild; 3433 WhereOrSet sSum, sCur; 3434 struct SrcList_item *pItem; 3435 3436 pWC = pBuilder->pWC; 3437 pWCEnd = pWC->a + pWC->nTerm; 3438 pNew = pBuilder->pNew; 3439 memset(&sSum, 0, sizeof(sSum)); 3440 pItem = pWInfo->pTabList->a + pNew->iTab; 3441 iCur = pItem->iCursor; 3442 3443 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3444 if( (pTerm->eOperator & WO_OR)!=0 3445 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3446 ){ 3447 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3448 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3449 WhereTerm *pOrTerm; 3450 int once = 1; 3451 int i, j; 3452 3453 sSubBuild = *pBuilder; 3454 sSubBuild.pOrderBy = 0; 3455 sSubBuild.pOrSet = &sCur; 3456 3457 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3458 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3459 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3460 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3461 }else if( pOrTerm->leftCursor==iCur ){ 3462 tempWC.pWInfo = pWC->pWInfo; 3463 tempWC.pOuter = pWC; 3464 tempWC.op = TK_AND; 3465 tempWC.nTerm = 1; 3466 tempWC.a = pOrTerm; 3467 sSubBuild.pWC = &tempWC; 3468 }else{ 3469 continue; 3470 } 3471 sCur.n = 0; 3472 #ifdef WHERETRACE_ENABLED 3473 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3474 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3475 if( sqlite3WhereTrace & 0x400 ){ 3476 sqlite3WhereClausePrint(sSubBuild.pWC); 3477 } 3478 #endif 3479 #ifndef SQLITE_OMIT_VIRTUALTABLE 3480 if( IsVirtual(pItem->pTab) ){ 3481 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 3482 }else 3483 #endif 3484 { 3485 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 3486 } 3487 if( rc==SQLITE_OK ){ 3488 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 3489 } 3490 assert( rc==SQLITE_OK || sCur.n==0 ); 3491 if( sCur.n==0 ){ 3492 sSum.n = 0; 3493 break; 3494 }else if( once ){ 3495 whereOrMove(&sSum, &sCur); 3496 once = 0; 3497 }else{ 3498 WhereOrSet sPrev; 3499 whereOrMove(&sPrev, &sSum); 3500 sSum.n = 0; 3501 for(i=0; i<sPrev.n; i++){ 3502 for(j=0; j<sCur.n; j++){ 3503 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3504 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3505 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3506 } 3507 } 3508 } 3509 } 3510 pNew->nLTerm = 1; 3511 pNew->aLTerm[0] = pTerm; 3512 pNew->wsFlags = WHERE_MULTI_OR; 3513 pNew->rSetup = 0; 3514 pNew->iSortIdx = 0; 3515 memset(&pNew->u, 0, sizeof(pNew->u)); 3516 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3517 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3518 ** of all sub-scans required by the OR-scan. However, due to rounding 3519 ** errors, it may be that the cost of the OR-scan is equal to its 3520 ** most expensive sub-scan. Add the smallest possible penalty 3521 ** (equivalent to multiplying the cost by 1.07) to ensure that 3522 ** this does not happen. Otherwise, for WHERE clauses such as the 3523 ** following where there is an index on "y": 3524 ** 3525 ** WHERE likelihood(x=?, 0.99) OR y=? 3526 ** 3527 ** the planner may elect to "OR" together a full-table scan and an 3528 ** index lookup. And other similarly odd results. */ 3529 pNew->rRun = sSum.a[i].rRun + 1; 3530 pNew->nOut = sSum.a[i].nOut; 3531 pNew->prereq = sSum.a[i].prereq; 3532 rc = whereLoopInsert(pBuilder, pNew); 3533 } 3534 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3535 } 3536 } 3537 return rc; 3538 } 3539 3540 /* 3541 ** Add all WhereLoop objects for all tables 3542 */ 3543 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3544 WhereInfo *pWInfo = pBuilder->pWInfo; 3545 Bitmask mPrereq = 0; 3546 Bitmask mPrior = 0; 3547 int iTab; 3548 SrcList *pTabList = pWInfo->pTabList; 3549 struct SrcList_item *pItem; 3550 struct SrcList_item *pEnd = &pTabList->a[pWInfo->nLevel]; 3551 sqlite3 *db = pWInfo->pParse->db; 3552 int rc = SQLITE_OK; 3553 WhereLoop *pNew; 3554 u8 priorJointype = 0; 3555 3556 /* Loop over the tables in the join, from left to right */ 3557 pNew = pBuilder->pNew; 3558 whereLoopInit(pNew); 3559 pBuilder->iPlanLimit = SQLITE_QUERY_PLANNER_LIMIT; 3560 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3561 Bitmask mUnusable = 0; 3562 pNew->iTab = iTab; 3563 pBuilder->iPlanLimit += SQLITE_QUERY_PLANNER_LIMIT_INCR; 3564 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3565 if( ((pItem->fg.jointype|priorJointype) & (JT_LEFT|JT_CROSS))!=0 ){ 3566 /* This condition is true when pItem is the FROM clause term on the 3567 ** right-hand-side of a LEFT or CROSS JOIN. */ 3568 mPrereq = mPrior; 3569 } 3570 priorJointype = pItem->fg.jointype; 3571 #ifndef SQLITE_OMIT_VIRTUALTABLE 3572 if( IsVirtual(pItem->pTab) ){ 3573 struct SrcList_item *p; 3574 for(p=&pItem[1]; p<pEnd; p++){ 3575 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3576 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3577 } 3578 } 3579 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 3580 }else 3581 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3582 { 3583 rc = whereLoopAddBtree(pBuilder, mPrereq); 3584 } 3585 if( rc==SQLITE_OK && pBuilder->pWC->hasOr ){ 3586 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 3587 } 3588 mPrior |= pNew->maskSelf; 3589 if( rc || db->mallocFailed ){ 3590 if( rc==SQLITE_DONE ){ 3591 /* We hit the query planner search limit set by iPlanLimit */ 3592 sqlite3_log(SQLITE_WARNING, "abbreviated query algorithm search"); 3593 rc = SQLITE_OK; 3594 }else{ 3595 break; 3596 } 3597 } 3598 } 3599 3600 whereLoopClear(db, pNew); 3601 return rc; 3602 } 3603 3604 /* 3605 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 3606 ** parameters) to see if it outputs rows in the requested ORDER BY 3607 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3608 ** 3609 ** N>0: N terms of the ORDER BY clause are satisfied 3610 ** N==0: No terms of the ORDER BY clause are satisfied 3611 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3612 ** 3613 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3614 ** strict. With GROUP BY and DISTINCT the only requirement is that 3615 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3616 ** and DISTINCT do not require rows to appear in any particular order as long 3617 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3618 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3619 ** pOrderBy terms must be matched in strict left-to-right order. 3620 */ 3621 static i8 wherePathSatisfiesOrderBy( 3622 WhereInfo *pWInfo, /* The WHERE clause */ 3623 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3624 WherePath *pPath, /* The WherePath to check */ 3625 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 3626 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3627 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3628 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3629 ){ 3630 u8 revSet; /* True if rev is known */ 3631 u8 rev; /* Composite sort order */ 3632 u8 revIdx; /* Index sort order */ 3633 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 3634 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 3635 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 3636 u16 eqOpMask; /* Allowed equality operators */ 3637 u16 nKeyCol; /* Number of key columns in pIndex */ 3638 u16 nColumn; /* Total number of ordered columns in the index */ 3639 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 3640 int iLoop; /* Index of WhereLoop in pPath being processed */ 3641 int i, j; /* Loop counters */ 3642 int iCur; /* Cursor number for current WhereLoop */ 3643 int iColumn; /* A column number within table iCur */ 3644 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 3645 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3646 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 3647 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 3648 Index *pIndex; /* The index associated with pLoop */ 3649 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 3650 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 3651 Bitmask obDone; /* Mask of all ORDER BY terms */ 3652 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 3653 Bitmask ready; /* Mask of inner loops */ 3654 3655 /* 3656 ** We say the WhereLoop is "one-row" if it generates no more than one 3657 ** row of output. A WhereLoop is one-row if all of the following are true: 3658 ** (a) All index columns match with WHERE_COLUMN_EQ. 3659 ** (b) The index is unique 3660 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 3661 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 3662 ** 3663 ** We say the WhereLoop is "order-distinct" if the set of columns from 3664 ** that WhereLoop that are in the ORDER BY clause are different for every 3665 ** row of the WhereLoop. Every one-row WhereLoop is automatically 3666 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 3667 ** is not order-distinct. To be order-distinct is not quite the same as being 3668 ** UNIQUE since a UNIQUE column or index can have multiple rows that 3669 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 3670 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 3671 ** 3672 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 3673 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 3674 ** automatically order-distinct. 3675 */ 3676 3677 assert( pOrderBy!=0 ); 3678 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 3679 3680 nOrderBy = pOrderBy->nExpr; 3681 testcase( nOrderBy==BMS-1 ); 3682 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 3683 isOrderDistinct = 1; 3684 obDone = MASKBIT(nOrderBy)-1; 3685 orderDistinctMask = 0; 3686 ready = 0; 3687 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 3688 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) eqOpMask |= WO_IN; 3689 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 3690 if( iLoop>0 ) ready |= pLoop->maskSelf; 3691 if( iLoop<nLoop ){ 3692 pLoop = pPath->aLoop[iLoop]; 3693 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 3694 }else{ 3695 pLoop = pLast; 3696 } 3697 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 3698 if( pLoop->u.vtab.isOrdered ) obSat = obDone; 3699 break; 3700 }else if( wctrlFlags & WHERE_DISTINCTBY ){ 3701 pLoop->u.btree.nDistinctCol = 0; 3702 } 3703 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 3704 3705 /* Mark off any ORDER BY term X that is a column in the table of 3706 ** the current loop for which there is term in the WHERE 3707 ** clause of the form X IS NULL or X=? that reference only outer 3708 ** loops. 3709 */ 3710 for(i=0; i<nOrderBy; i++){ 3711 if( MASKBIT(i) & obSat ) continue; 3712 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3713 if( pOBExpr->op!=TK_COLUMN ) continue; 3714 if( pOBExpr->iTable!=iCur ) continue; 3715 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 3716 ~ready, eqOpMask, 0); 3717 if( pTerm==0 ) continue; 3718 if( pTerm->eOperator==WO_IN ){ 3719 /* IN terms are only valid for sorting in the ORDER BY LIMIT 3720 ** optimization, and then only if they are actually used 3721 ** by the query plan */ 3722 assert( wctrlFlags & WHERE_ORDERBY_LIMIT ); 3723 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 3724 if( j>=pLoop->nLTerm ) continue; 3725 } 3726 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 3727 if( sqlite3ExprCollSeqMatch(pWInfo->pParse, 3728 pOrderBy->a[i].pExpr, pTerm->pExpr)==0 ){ 3729 continue; 3730 } 3731 testcase( pTerm->pExpr->op==TK_IS ); 3732 } 3733 obSat |= MASKBIT(i); 3734 } 3735 3736 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 3737 if( pLoop->wsFlags & WHERE_IPK ){ 3738 pIndex = 0; 3739 nKeyCol = 0; 3740 nColumn = 1; 3741 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 3742 return 0; 3743 }else{ 3744 nKeyCol = pIndex->nKeyCol; 3745 nColumn = pIndex->nColumn; 3746 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 3747 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 3748 || !HasRowid(pIndex->pTable)); 3749 isOrderDistinct = IsUniqueIndex(pIndex) 3750 && (pLoop->wsFlags & WHERE_SKIPSCAN)==0; 3751 } 3752 3753 /* Loop through all columns of the index and deal with the ones 3754 ** that are not constrained by == or IN. 3755 */ 3756 rev = revSet = 0; 3757 distinctColumns = 0; 3758 for(j=0; j<nColumn; j++){ 3759 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 3760 3761 assert( j>=pLoop->u.btree.nEq 3762 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 3763 ); 3764 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 3765 u16 eOp = pLoop->aLTerm[j]->eOperator; 3766 3767 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 3768 ** doing WHERE_ORDERBY_LIMIT processing). 3769 ** 3770 ** If the current term is a column of an ((?,?) IN (SELECT...)) 3771 ** expression for which the SELECT returns more than one column, 3772 ** check that it is the only column used by this loop. Otherwise, 3773 ** if it is one of two or more, none of the columns can be 3774 ** considered to match an ORDER BY term. */ 3775 if( (eOp & eqOpMask)!=0 ){ 3776 if( eOp & WO_ISNULL ){ 3777 testcase( isOrderDistinct ); 3778 isOrderDistinct = 0; 3779 } 3780 continue; 3781 }else if( ALWAYS(eOp & WO_IN) ){ 3782 /* ALWAYS() justification: eOp is an equality operator due to the 3783 ** j<pLoop->u.btree.nEq constraint above. Any equality other 3784 ** than WO_IN is captured by the previous "if". So this one 3785 ** always has to be WO_IN. */ 3786 Expr *pX = pLoop->aLTerm[j]->pExpr; 3787 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 3788 if( pLoop->aLTerm[i]->pExpr==pX ){ 3789 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 3790 bOnce = 0; 3791 break; 3792 } 3793 } 3794 } 3795 } 3796 3797 /* Get the column number in the table (iColumn) and sort order 3798 ** (revIdx) for the j-th column of the index. 3799 */ 3800 if( pIndex ){ 3801 iColumn = pIndex->aiColumn[j]; 3802 revIdx = pIndex->aSortOrder[j]; 3803 if( iColumn==pIndex->pTable->iPKey ) iColumn = XN_ROWID; 3804 }else{ 3805 iColumn = XN_ROWID; 3806 revIdx = 0; 3807 } 3808 3809 /* An unconstrained column that might be NULL means that this 3810 ** WhereLoop is not well-ordered 3811 */ 3812 if( isOrderDistinct 3813 && iColumn>=0 3814 && j>=pLoop->u.btree.nEq 3815 && pIndex->pTable->aCol[iColumn].notNull==0 3816 ){ 3817 isOrderDistinct = 0; 3818 } 3819 3820 /* Find the ORDER BY term that corresponds to the j-th column 3821 ** of the index and mark that ORDER BY term off 3822 */ 3823 isMatch = 0; 3824 for(i=0; bOnce && i<nOrderBy; i++){ 3825 if( MASKBIT(i) & obSat ) continue; 3826 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3827 testcase( wctrlFlags & WHERE_GROUPBY ); 3828 testcase( wctrlFlags & WHERE_DISTINCTBY ); 3829 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 3830 if( iColumn>=XN_ROWID ){ 3831 if( pOBExpr->op!=TK_COLUMN ) continue; 3832 if( pOBExpr->iTable!=iCur ) continue; 3833 if( pOBExpr->iColumn!=iColumn ) continue; 3834 }else{ 3835 Expr *pIdxExpr = pIndex->aColExpr->a[j].pExpr; 3836 if( sqlite3ExprCompareSkip(pOBExpr, pIdxExpr, iCur) ){ 3837 continue; 3838 } 3839 } 3840 if( iColumn!=XN_ROWID ){ 3841 pColl = sqlite3ExprNNCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3842 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 3843 } 3844 if( wctrlFlags & WHERE_DISTINCTBY ){ 3845 pLoop->u.btree.nDistinctCol = j+1; 3846 } 3847 isMatch = 1; 3848 break; 3849 } 3850 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 3851 /* Make sure the sort order is compatible in an ORDER BY clause. 3852 ** Sort order is irrelevant for a GROUP BY clause. */ 3853 if( revSet ){ 3854 if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0; 3855 }else{ 3856 rev = revIdx ^ pOrderBy->a[i].sortOrder; 3857 if( rev ) *pRevMask |= MASKBIT(iLoop); 3858 revSet = 1; 3859 } 3860 } 3861 if( isMatch ){ 3862 if( iColumn==XN_ROWID ){ 3863 testcase( distinctColumns==0 ); 3864 distinctColumns = 1; 3865 } 3866 obSat |= MASKBIT(i); 3867 if( (wctrlFlags & WHERE_ORDERBY_MIN) && j==pLoop->u.btree.nEq ){ 3868 pLoop->wsFlags |= WHERE_MIN_ORDERED; 3869 } 3870 }else{ 3871 /* No match found */ 3872 if( j==0 || j<nKeyCol ){ 3873 testcase( isOrderDistinct!=0 ); 3874 isOrderDistinct = 0; 3875 } 3876 break; 3877 } 3878 } /* end Loop over all index columns */ 3879 if( distinctColumns ){ 3880 testcase( isOrderDistinct==0 ); 3881 isOrderDistinct = 1; 3882 } 3883 } /* end-if not one-row */ 3884 3885 /* Mark off any other ORDER BY terms that reference pLoop */ 3886 if( isOrderDistinct ){ 3887 orderDistinctMask |= pLoop->maskSelf; 3888 for(i=0; i<nOrderBy; i++){ 3889 Expr *p; 3890 Bitmask mTerm; 3891 if( MASKBIT(i) & obSat ) continue; 3892 p = pOrderBy->a[i].pExpr; 3893 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 3894 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 3895 if( (mTerm&~orderDistinctMask)==0 ){ 3896 obSat |= MASKBIT(i); 3897 } 3898 } 3899 } 3900 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 3901 if( obSat==obDone ) return (i8)nOrderBy; 3902 if( !isOrderDistinct ){ 3903 for(i=nOrderBy-1; i>0; i--){ 3904 Bitmask m = MASKBIT(i) - 1; 3905 if( (obSat&m)==m ) return i; 3906 } 3907 return 0; 3908 } 3909 return -1; 3910 } 3911 3912 3913 /* 3914 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 3915 ** the planner assumes that the specified pOrderBy list is actually a GROUP 3916 ** BY clause - and so any order that groups rows as required satisfies the 3917 ** request. 3918 ** 3919 ** Normally, in this case it is not possible for the caller to determine 3920 ** whether or not the rows are really being delivered in sorted order, or 3921 ** just in some other order that provides the required grouping. However, 3922 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 3923 ** this function may be called on the returned WhereInfo object. It returns 3924 ** true if the rows really will be sorted in the specified order, or false 3925 ** otherwise. 3926 ** 3927 ** For example, assuming: 3928 ** 3929 ** CREATE INDEX i1 ON t1(x, Y); 3930 ** 3931 ** then 3932 ** 3933 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 3934 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 3935 */ 3936 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 3937 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 3938 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 3939 return pWInfo->sorted; 3940 } 3941 3942 #ifdef WHERETRACE_ENABLED 3943 /* For debugging use only: */ 3944 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 3945 static char zName[65]; 3946 int i; 3947 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 3948 if( pLast ) zName[i++] = pLast->cId; 3949 zName[i] = 0; 3950 return zName; 3951 } 3952 #endif 3953 3954 /* 3955 ** Return the cost of sorting nRow rows, assuming that the keys have 3956 ** nOrderby columns and that the first nSorted columns are already in 3957 ** order. 3958 */ 3959 static LogEst whereSortingCost( 3960 WhereInfo *pWInfo, 3961 LogEst nRow, 3962 int nOrderBy, 3963 int nSorted 3964 ){ 3965 /* TUNING: Estimated cost of a full external sort, where N is 3966 ** the number of rows to sort is: 3967 ** 3968 ** cost = (3.0 * N * log(N)). 3969 ** 3970 ** Or, if the order-by clause has X terms but only the last Y 3971 ** terms are out of order, then block-sorting will reduce the 3972 ** sorting cost to: 3973 ** 3974 ** cost = (3.0 * N * log(N)) * (Y/X) 3975 ** 3976 ** The (Y/X) term is implemented using stack variable rScale 3977 ** below. */ 3978 LogEst rScale, rSortCost; 3979 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 3980 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 3981 rSortCost = nRow + rScale + 16; 3982 3983 /* Multiple by log(M) where M is the number of output rows. 3984 ** Use the LIMIT for M if it is smaller */ 3985 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 3986 nRow = pWInfo->iLimit; 3987 } 3988 rSortCost += estLog(nRow); 3989 return rSortCost; 3990 } 3991 3992 /* 3993 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 3994 ** attempts to find the lowest cost path that visits each WhereLoop 3995 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 3996 ** 3997 ** Assume that the total number of output rows that will need to be sorted 3998 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 3999 ** costs if nRowEst==0. 4000 ** 4001 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 4002 ** error occurs. 4003 */ 4004 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 4005 int mxChoice; /* Maximum number of simultaneous paths tracked */ 4006 int nLoop; /* Number of terms in the join */ 4007 Parse *pParse; /* Parsing context */ 4008 sqlite3 *db; /* The database connection */ 4009 int iLoop; /* Loop counter over the terms of the join */ 4010 int ii, jj; /* Loop counters */ 4011 int mxI = 0; /* Index of next entry to replace */ 4012 int nOrderBy; /* Number of ORDER BY clause terms */ 4013 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 4014 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 4015 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 4016 WherePath *aFrom; /* All nFrom paths at the previous level */ 4017 WherePath *aTo; /* The nTo best paths at the current level */ 4018 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 4019 WherePath *pTo; /* An element of aTo[] that we are working on */ 4020 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 4021 WhereLoop **pX; /* Used to divy up the pSpace memory */ 4022 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 4023 char *pSpace; /* Temporary memory used by this routine */ 4024 int nSpace; /* Bytes of space allocated at pSpace */ 4025 4026 pParse = pWInfo->pParse; 4027 db = pParse->db; 4028 nLoop = pWInfo->nLevel; 4029 /* TUNING: For simple queries, only the best path is tracked. 4030 ** For 2-way joins, the 5 best paths are followed. 4031 ** For joins of 3 or more tables, track the 10 best paths */ 4032 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 4033 assert( nLoop<=pWInfo->pTabList->nSrc ); 4034 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 4035 4036 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 4037 ** case the purpose of this call is to estimate the number of rows returned 4038 ** by the overall query. Once this estimate has been obtained, the caller 4039 ** will invoke this function a second time, passing the estimate as the 4040 ** nRowEst parameter. */ 4041 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 4042 nOrderBy = 0; 4043 }else{ 4044 nOrderBy = pWInfo->pOrderBy->nExpr; 4045 } 4046 4047 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 4048 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 4049 nSpace += sizeof(LogEst) * nOrderBy; 4050 pSpace = sqlite3DbMallocRawNN(db, nSpace); 4051 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 4052 aTo = (WherePath*)pSpace; 4053 aFrom = aTo+mxChoice; 4054 memset(aFrom, 0, sizeof(aFrom[0])); 4055 pX = (WhereLoop**)(aFrom+mxChoice); 4056 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 4057 pFrom->aLoop = pX; 4058 } 4059 if( nOrderBy ){ 4060 /* If there is an ORDER BY clause and it is not being ignored, set up 4061 ** space for the aSortCost[] array. Each element of the aSortCost array 4062 ** is either zero - meaning it has not yet been initialized - or the 4063 ** cost of sorting nRowEst rows of data where the first X terms of 4064 ** the ORDER BY clause are already in order, where X is the array 4065 ** index. */ 4066 aSortCost = (LogEst*)pX; 4067 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 4068 } 4069 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 4070 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 4071 4072 /* Seed the search with a single WherePath containing zero WhereLoops. 4073 ** 4074 ** TUNING: Do not let the number of iterations go above 28. If the cost 4075 ** of computing an automatic index is not paid back within the first 28 4076 ** rows, then do not use the automatic index. */ 4077 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 4078 nFrom = 1; 4079 assert( aFrom[0].isOrdered==0 ); 4080 if( nOrderBy ){ 4081 /* If nLoop is zero, then there are no FROM terms in the query. Since 4082 ** in this case the query may return a maximum of one row, the results 4083 ** are already in the requested order. Set isOrdered to nOrderBy to 4084 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 4085 ** -1, indicating that the result set may or may not be ordered, 4086 ** depending on the loops added to the current plan. */ 4087 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 4088 } 4089 4090 /* Compute successively longer WherePaths using the previous generation 4091 ** of WherePaths as the basis for the next. Keep track of the mxChoice 4092 ** best paths at each generation */ 4093 for(iLoop=0; iLoop<nLoop; iLoop++){ 4094 nTo = 0; 4095 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 4096 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 4097 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 4098 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 4099 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 4100 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 4101 Bitmask maskNew; /* Mask of src visited by (..) */ 4102 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 4103 4104 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 4105 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 4106 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<3 ){ 4107 /* Do not use an automatic index if the this loop is expected 4108 ** to run less than 1.25 times. It is tempting to also exclude 4109 ** automatic index usage on an outer loop, but sometimes an automatic 4110 ** index is useful in the outer loop of a correlated subquery. */ 4111 assert( 10==sqlite3LogEst(2) ); 4112 continue; 4113 } 4114 4115 /* At this point, pWLoop is a candidate to be the next loop. 4116 ** Compute its cost */ 4117 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 4118 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 4119 nOut = pFrom->nRow + pWLoop->nOut; 4120 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 4121 if( isOrdered<0 ){ 4122 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 4123 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 4124 iLoop, pWLoop, &revMask); 4125 }else{ 4126 revMask = pFrom->revLoop; 4127 } 4128 if( isOrdered>=0 && isOrdered<nOrderBy ){ 4129 if( aSortCost[isOrdered]==0 ){ 4130 aSortCost[isOrdered] = whereSortingCost( 4131 pWInfo, nRowEst, nOrderBy, isOrdered 4132 ); 4133 } 4134 /* TUNING: Add a small extra penalty (5) to sorting as an 4135 ** extra encouragment to the query planner to select a plan 4136 ** where the rows emerge in the correct order without any sorting 4137 ** required. */ 4138 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]) + 5; 4139 4140 WHERETRACE(0x002, 4141 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 4142 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 4143 rUnsorted, rCost)); 4144 }else{ 4145 rCost = rUnsorted; 4146 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 4147 } 4148 4149 /* Check to see if pWLoop should be added to the set of 4150 ** mxChoice best-so-far paths. 4151 ** 4152 ** First look for an existing path among best-so-far paths 4153 ** that covers the same set of loops and has the same isOrdered 4154 ** setting as the current path candidate. 4155 ** 4156 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 4157 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 4158 ** of legal values for isOrdered, -1..64. 4159 */ 4160 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4161 if( pTo->maskLoop==maskNew 4162 && ((pTo->isOrdered^isOrdered)&0x80)==0 4163 ){ 4164 testcase( jj==nTo-1 ); 4165 break; 4166 } 4167 } 4168 if( jj>=nTo ){ 4169 /* None of the existing best-so-far paths match the candidate. */ 4170 if( nTo>=mxChoice 4171 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4172 ){ 4173 /* The current candidate is no better than any of the mxChoice 4174 ** paths currently in the best-so-far buffer. So discard 4175 ** this candidate as not viable. */ 4176 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4177 if( sqlite3WhereTrace&0x4 ){ 4178 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4179 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4180 isOrdered>=0 ? isOrdered+'0' : '?'); 4181 } 4182 #endif 4183 continue; 4184 } 4185 /* If we reach this points it means that the new candidate path 4186 ** needs to be added to the set of best-so-far paths. */ 4187 if( nTo<mxChoice ){ 4188 /* Increase the size of the aTo set by one */ 4189 jj = nTo++; 4190 }else{ 4191 /* New path replaces the prior worst to keep count below mxChoice */ 4192 jj = mxI; 4193 } 4194 pTo = &aTo[jj]; 4195 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4196 if( sqlite3WhereTrace&0x4 ){ 4197 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4198 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4199 isOrdered>=0 ? isOrdered+'0' : '?'); 4200 } 4201 #endif 4202 }else{ 4203 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4204 ** same set of loops and has the same isOrdered setting as the 4205 ** candidate path. Check to see if the candidate should replace 4206 ** pTo or if the candidate should be skipped. 4207 ** 4208 ** The conditional is an expanded vector comparison equivalent to: 4209 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4210 */ 4211 if( pTo->rCost<rCost 4212 || (pTo->rCost==rCost 4213 && (pTo->nRow<nOut 4214 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4215 ) 4216 ) 4217 ){ 4218 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4219 if( sqlite3WhereTrace&0x4 ){ 4220 sqlite3DebugPrintf( 4221 "Skip %s cost=%-3d,%3d,%3d order=%c", 4222 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4223 isOrdered>=0 ? isOrdered+'0' : '?'); 4224 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4225 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4226 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4227 } 4228 #endif 4229 /* Discard the candidate path from further consideration */ 4230 testcase( pTo->rCost==rCost ); 4231 continue; 4232 } 4233 testcase( pTo->rCost==rCost+1 ); 4234 /* Control reaches here if the candidate path is better than the 4235 ** pTo path. Replace pTo with the candidate. */ 4236 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4237 if( sqlite3WhereTrace&0x4 ){ 4238 sqlite3DebugPrintf( 4239 "Update %s cost=%-3d,%3d,%3d order=%c", 4240 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4241 isOrdered>=0 ? isOrdered+'0' : '?'); 4242 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4243 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4244 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4245 } 4246 #endif 4247 } 4248 /* pWLoop is a winner. Add it to the set of best so far */ 4249 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4250 pTo->revLoop = revMask; 4251 pTo->nRow = nOut; 4252 pTo->rCost = rCost; 4253 pTo->rUnsorted = rUnsorted; 4254 pTo->isOrdered = isOrdered; 4255 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4256 pTo->aLoop[iLoop] = pWLoop; 4257 if( nTo>=mxChoice ){ 4258 mxI = 0; 4259 mxCost = aTo[0].rCost; 4260 mxUnsorted = aTo[0].nRow; 4261 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4262 if( pTo->rCost>mxCost 4263 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4264 ){ 4265 mxCost = pTo->rCost; 4266 mxUnsorted = pTo->rUnsorted; 4267 mxI = jj; 4268 } 4269 } 4270 } 4271 } 4272 } 4273 4274 #ifdef WHERETRACE_ENABLED /* >=2 */ 4275 if( sqlite3WhereTrace & 0x02 ){ 4276 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4277 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4278 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4279 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4280 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4281 if( pTo->isOrdered>0 ){ 4282 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4283 }else{ 4284 sqlite3DebugPrintf("\n"); 4285 } 4286 } 4287 } 4288 #endif 4289 4290 /* Swap the roles of aFrom and aTo for the next generation */ 4291 pFrom = aTo; 4292 aTo = aFrom; 4293 aFrom = pFrom; 4294 nFrom = nTo; 4295 } 4296 4297 if( nFrom==0 ){ 4298 sqlite3ErrorMsg(pParse, "no query solution"); 4299 sqlite3DbFreeNN(db, pSpace); 4300 return SQLITE_ERROR; 4301 } 4302 4303 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4304 pFrom = aFrom; 4305 for(ii=1; ii<nFrom; ii++){ 4306 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4307 } 4308 assert( pWInfo->nLevel==nLoop ); 4309 /* Load the lowest cost path into pWInfo */ 4310 for(iLoop=0; iLoop<nLoop; iLoop++){ 4311 WhereLevel *pLevel = pWInfo->a + iLoop; 4312 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4313 pLevel->iFrom = pWLoop->iTab; 4314 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4315 } 4316 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4317 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4318 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4319 && nRowEst 4320 ){ 4321 Bitmask notUsed; 4322 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4323 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4324 if( rc==pWInfo->pResultSet->nExpr ){ 4325 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4326 } 4327 } 4328 pWInfo->bOrderedInnerLoop = 0; 4329 if( pWInfo->pOrderBy ){ 4330 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4331 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4332 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4333 } 4334 }else{ 4335 pWInfo->nOBSat = pFrom->isOrdered; 4336 pWInfo->revMask = pFrom->revLoop; 4337 if( pWInfo->nOBSat<=0 ){ 4338 pWInfo->nOBSat = 0; 4339 if( nLoop>0 ){ 4340 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4341 if( (wsFlags & WHERE_ONEROW)==0 4342 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4343 ){ 4344 Bitmask m = 0; 4345 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4346 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4347 testcase( wsFlags & WHERE_IPK ); 4348 testcase( wsFlags & WHERE_COLUMN_IN ); 4349 if( rc==pWInfo->pOrderBy->nExpr ){ 4350 pWInfo->bOrderedInnerLoop = 1; 4351 pWInfo->revMask = m; 4352 } 4353 } 4354 } 4355 } 4356 } 4357 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4358 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4359 ){ 4360 Bitmask revMask = 0; 4361 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4362 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4363 ); 4364 assert( pWInfo->sorted==0 ); 4365 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4366 pWInfo->sorted = 1; 4367 pWInfo->revMask = revMask; 4368 } 4369 } 4370 } 4371 4372 4373 pWInfo->nRowOut = pFrom->nRow; 4374 4375 /* Free temporary memory and return success */ 4376 sqlite3DbFreeNN(db, pSpace); 4377 return SQLITE_OK; 4378 } 4379 4380 /* 4381 ** Most queries use only a single table (they are not joins) and have 4382 ** simple == constraints against indexed fields. This routine attempts 4383 ** to plan those simple cases using much less ceremony than the 4384 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4385 ** times for the common case. 4386 ** 4387 ** Return non-zero on success, if this query can be handled by this 4388 ** no-frills query planner. Return zero if this query needs the 4389 ** general-purpose query planner. 4390 */ 4391 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4392 WhereInfo *pWInfo; 4393 struct SrcList_item *pItem; 4394 WhereClause *pWC; 4395 WhereTerm *pTerm; 4396 WhereLoop *pLoop; 4397 int iCur; 4398 int j; 4399 Table *pTab; 4400 Index *pIdx; 4401 4402 pWInfo = pBuilder->pWInfo; 4403 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4404 assert( pWInfo->pTabList->nSrc>=1 ); 4405 pItem = pWInfo->pTabList->a; 4406 pTab = pItem->pTab; 4407 if( IsVirtual(pTab) ) return 0; 4408 if( pItem->fg.isIndexedBy ) return 0; 4409 iCur = pItem->iCursor; 4410 pWC = &pWInfo->sWC; 4411 pLoop = pBuilder->pNew; 4412 pLoop->wsFlags = 0; 4413 pLoop->nSkip = 0; 4414 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0); 4415 if( pTerm ){ 4416 testcase( pTerm->eOperator & WO_IS ); 4417 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4418 pLoop->aLTerm[0] = pTerm; 4419 pLoop->nLTerm = 1; 4420 pLoop->u.btree.nEq = 1; 4421 /* TUNING: Cost of a rowid lookup is 10 */ 4422 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 4423 }else{ 4424 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 4425 int opMask; 4426 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 4427 if( !IsUniqueIndex(pIdx) 4428 || pIdx->pPartIdxWhere!=0 4429 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 4430 ) continue; 4431 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 4432 for(j=0; j<pIdx->nKeyCol; j++){ 4433 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx); 4434 if( pTerm==0 ) break; 4435 testcase( pTerm->eOperator & WO_IS ); 4436 pLoop->aLTerm[j] = pTerm; 4437 } 4438 if( j!=pIdx->nKeyCol ) continue; 4439 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 4440 if( pIdx->isCovering || (pItem->colUsed & pIdx->colNotIdxed)==0 ){ 4441 pLoop->wsFlags |= WHERE_IDX_ONLY; 4442 } 4443 pLoop->nLTerm = j; 4444 pLoop->u.btree.nEq = j; 4445 pLoop->u.btree.pIndex = pIdx; 4446 /* TUNING: Cost of a unique index lookup is 15 */ 4447 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 4448 break; 4449 } 4450 } 4451 if( pLoop->wsFlags ){ 4452 pLoop->nOut = (LogEst)1; 4453 pWInfo->a[0].pWLoop = pLoop; 4454 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 4455 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 4456 pWInfo->a[0].iTabCur = iCur; 4457 pWInfo->nRowOut = 1; 4458 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 4459 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 4460 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4461 } 4462 #ifdef SQLITE_DEBUG 4463 pLoop->cId = '0'; 4464 #endif 4465 return 1; 4466 } 4467 return 0; 4468 } 4469 4470 /* 4471 ** Helper function for exprIsDeterministic(). 4472 */ 4473 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 4474 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 4475 pWalker->eCode = 0; 4476 return WRC_Abort; 4477 } 4478 return WRC_Continue; 4479 } 4480 4481 /* 4482 ** Return true if the expression contains no non-deterministic SQL 4483 ** functions. Do not consider non-deterministic SQL functions that are 4484 ** part of sub-select statements. 4485 */ 4486 static int exprIsDeterministic(Expr *p){ 4487 Walker w; 4488 memset(&w, 0, sizeof(w)); 4489 w.eCode = 1; 4490 w.xExprCallback = exprNodeIsDeterministic; 4491 w.xSelectCallback = sqlite3SelectWalkFail; 4492 sqlite3WalkExpr(&w, p); 4493 return w.eCode; 4494 } 4495 4496 /* 4497 ** Generate the beginning of the loop used for WHERE clause processing. 4498 ** The return value is a pointer to an opaque structure that contains 4499 ** information needed to terminate the loop. Later, the calling routine 4500 ** should invoke sqlite3WhereEnd() with the return value of this function 4501 ** in order to complete the WHERE clause processing. 4502 ** 4503 ** If an error occurs, this routine returns NULL. 4504 ** 4505 ** The basic idea is to do a nested loop, one loop for each table in 4506 ** the FROM clause of a select. (INSERT and UPDATE statements are the 4507 ** same as a SELECT with only a single table in the FROM clause.) For 4508 ** example, if the SQL is this: 4509 ** 4510 ** SELECT * FROM t1, t2, t3 WHERE ...; 4511 ** 4512 ** Then the code generated is conceptually like the following: 4513 ** 4514 ** foreach row1 in t1 do \ Code generated 4515 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 4516 ** foreach row3 in t3 do / 4517 ** ... 4518 ** end \ Code generated 4519 ** end |-- by sqlite3WhereEnd() 4520 ** end / 4521 ** 4522 ** Note that the loops might not be nested in the order in which they 4523 ** appear in the FROM clause if a different order is better able to make 4524 ** use of indices. Note also that when the IN operator appears in 4525 ** the WHERE clause, it might result in additional nested loops for 4526 ** scanning through all values on the right-hand side of the IN. 4527 ** 4528 ** There are Btree cursors associated with each table. t1 uses cursor 4529 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 4530 ** And so forth. This routine generates code to open those VDBE cursors 4531 ** and sqlite3WhereEnd() generates the code to close them. 4532 ** 4533 ** The code that sqlite3WhereBegin() generates leaves the cursors named 4534 ** in pTabList pointing at their appropriate entries. The [...] code 4535 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 4536 ** data from the various tables of the loop. 4537 ** 4538 ** If the WHERE clause is empty, the foreach loops must each scan their 4539 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 4540 ** the tables have indices and there are terms in the WHERE clause that 4541 ** refer to those indices, a complete table scan can be avoided and the 4542 ** code will run much faster. Most of the work of this routine is checking 4543 ** to see if there are indices that can be used to speed up the loop. 4544 ** 4545 ** Terms of the WHERE clause are also used to limit which rows actually 4546 ** make it to the "..." in the middle of the loop. After each "foreach", 4547 ** terms of the WHERE clause that use only terms in that loop and outer 4548 ** loops are evaluated and if false a jump is made around all subsequent 4549 ** inner loops (or around the "..." if the test occurs within the inner- 4550 ** most loop) 4551 ** 4552 ** OUTER JOINS 4553 ** 4554 ** An outer join of tables t1 and t2 is conceptally coded as follows: 4555 ** 4556 ** foreach row1 in t1 do 4557 ** flag = 0 4558 ** foreach row2 in t2 do 4559 ** start: 4560 ** ... 4561 ** flag = 1 4562 ** end 4563 ** if flag==0 then 4564 ** move the row2 cursor to a null row 4565 ** goto start 4566 ** fi 4567 ** end 4568 ** 4569 ** ORDER BY CLAUSE PROCESSING 4570 ** 4571 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 4572 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 4573 ** if there is one. If there is no ORDER BY clause or if this routine 4574 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 4575 ** 4576 ** The iIdxCur parameter is the cursor number of an index. If 4577 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 4578 ** to use for OR clause processing. The WHERE clause should use this 4579 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 4580 ** the first cursor in an array of cursors for all indices. iIdxCur should 4581 ** be used to compute the appropriate cursor depending on which index is 4582 ** used. 4583 */ 4584 WhereInfo *sqlite3WhereBegin( 4585 Parse *pParse, /* The parser context */ 4586 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 4587 Expr *pWhere, /* The WHERE clause */ 4588 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 4589 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 4590 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 4591 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 4592 ** If WHERE_USE_LIMIT, then the limit amount */ 4593 ){ 4594 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4595 int nTabList; /* Number of elements in pTabList */ 4596 WhereInfo *pWInfo; /* Will become the return value of this function */ 4597 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4598 Bitmask notReady; /* Cursors that are not yet positioned */ 4599 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 4600 WhereMaskSet *pMaskSet; /* The expression mask set */ 4601 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 4602 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 4603 int ii; /* Loop counter */ 4604 sqlite3 *db; /* Database connection */ 4605 int rc; /* Return code */ 4606 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 4607 4608 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 4609 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 4610 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4611 )); 4612 4613 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 4614 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4615 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 4616 4617 /* Variable initialization */ 4618 db = pParse->db; 4619 memset(&sWLB, 0, sizeof(sWLB)); 4620 4621 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 4622 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 4623 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 4624 sWLB.pOrderBy = pOrderBy; 4625 4626 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 4627 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 4628 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 4629 wctrlFlags &= ~WHERE_WANT_DISTINCT; 4630 } 4631 4632 /* The number of tables in the FROM clause is limited by the number of 4633 ** bits in a Bitmask 4634 */ 4635 testcase( pTabList->nSrc==BMS ); 4636 if( pTabList->nSrc>BMS ){ 4637 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 4638 return 0; 4639 } 4640 4641 /* This function normally generates a nested loop for all tables in 4642 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 4643 ** only generate code for the first table in pTabList and assume that 4644 ** any cursors associated with subsequent tables are uninitialized. 4645 */ 4646 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 4647 4648 /* Allocate and initialize the WhereInfo structure that will become the 4649 ** return value. A single allocation is used to store the WhereInfo 4650 ** struct, the contents of WhereInfo.a[], the WhereClause structure 4651 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 4652 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 4653 ** some architectures. Hence the ROUND8() below. 4654 */ 4655 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 4656 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 4657 if( db->mallocFailed ){ 4658 sqlite3DbFree(db, pWInfo); 4659 pWInfo = 0; 4660 goto whereBeginError; 4661 } 4662 pWInfo->pParse = pParse; 4663 pWInfo->pTabList = pTabList; 4664 pWInfo->pOrderBy = pOrderBy; 4665 pWInfo->pWhere = pWhere; 4666 pWInfo->pResultSet = pResultSet; 4667 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 4668 pWInfo->nLevel = nTabList; 4669 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(pParse); 4670 pWInfo->wctrlFlags = wctrlFlags; 4671 pWInfo->iLimit = iAuxArg; 4672 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 4673 memset(&pWInfo->nOBSat, 0, 4674 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 4675 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 4676 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 4677 pMaskSet = &pWInfo->sMaskSet; 4678 sWLB.pWInfo = pWInfo; 4679 sWLB.pWC = &pWInfo->sWC; 4680 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 4681 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 4682 whereLoopInit(sWLB.pNew); 4683 #ifdef SQLITE_DEBUG 4684 sWLB.pNew->cId = '*'; 4685 #endif 4686 4687 /* Split the WHERE clause into separate subexpressions where each 4688 ** subexpression is separated by an AND operator. 4689 */ 4690 initMaskSet(pMaskSet); 4691 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 4692 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 4693 4694 /* Special case: No FROM clause 4695 */ 4696 if( nTabList==0 ){ 4697 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 4698 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4699 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4700 } 4701 ExplainQueryPlan((pParse, 0, "SCAN CONSTANT ROW")); 4702 }else{ 4703 /* Assign a bit from the bitmask to every term in the FROM clause. 4704 ** 4705 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 4706 ** 4707 ** The rule of the previous sentence ensures thta if X is the bitmask for 4708 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 4709 ** Knowing the bitmask for all tables to the left of a left join is 4710 ** important. Ticket #3015. 4711 ** 4712 ** Note that bitmasks are created for all pTabList->nSrc tables in 4713 ** pTabList, not just the first nTabList tables. nTabList is normally 4714 ** equal to pTabList->nSrc but might be shortened to 1 if the 4715 ** WHERE_OR_SUBCLAUSE flag is set. 4716 */ 4717 ii = 0; 4718 do{ 4719 createMask(pMaskSet, pTabList->a[ii].iCursor); 4720 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 4721 }while( (++ii)<pTabList->nSrc ); 4722 #ifdef SQLITE_DEBUG 4723 { 4724 Bitmask mx = 0; 4725 for(ii=0; ii<pTabList->nSrc; ii++){ 4726 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 4727 assert( m>=mx ); 4728 mx = m; 4729 } 4730 } 4731 #endif 4732 } 4733 4734 /* Analyze all of the subexpressions. */ 4735 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 4736 if( db->mallocFailed ) goto whereBeginError; 4737 4738 /* Special case: WHERE terms that do not refer to any tables in the join 4739 ** (constant expressions). Evaluate each such term, and jump over all the 4740 ** generated code if the result is not true. 4741 ** 4742 ** Do not do this if the expression contains non-deterministic functions 4743 ** that are not within a sub-select. This is not strictly required, but 4744 ** preserves SQLite's legacy behaviour in the following two cases: 4745 ** 4746 ** FROM ... WHERE random()>0; -- eval random() once per row 4747 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 4748 */ 4749 for(ii=0; ii<sWLB.pWC->nTerm; ii++){ 4750 WhereTerm *pT = &sWLB.pWC->a[ii]; 4751 if( pT->wtFlags & TERM_VIRTUAL ) continue; 4752 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 4753 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 4754 pT->wtFlags |= TERM_CODED; 4755 } 4756 } 4757 4758 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4759 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 4760 /* The DISTINCT marking is pointless. Ignore it. */ 4761 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4762 }else if( pOrderBy==0 ){ 4763 /* Try to ORDER BY the result set to make distinct processing easier */ 4764 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 4765 pWInfo->pOrderBy = pResultSet; 4766 } 4767 } 4768 4769 /* Construct the WhereLoop objects */ 4770 #if defined(WHERETRACE_ENABLED) 4771 if( sqlite3WhereTrace & 0xffff ){ 4772 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 4773 if( wctrlFlags & WHERE_USE_LIMIT ){ 4774 sqlite3DebugPrintf(", limit: %d", iAuxArg); 4775 } 4776 sqlite3DebugPrintf(")\n"); 4777 if( sqlite3WhereTrace & 0x100 ){ 4778 Select sSelect; 4779 memset(&sSelect, 0, sizeof(sSelect)); 4780 sSelect.selFlags = SF_WhereBegin; 4781 sSelect.pSrc = pTabList; 4782 sSelect.pWhere = pWhere; 4783 sSelect.pOrderBy = pOrderBy; 4784 sSelect.pEList = pResultSet; 4785 sqlite3TreeViewSelect(0, &sSelect, 0); 4786 } 4787 } 4788 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 4789 sqlite3WhereClausePrint(sWLB.pWC); 4790 } 4791 #endif 4792 4793 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 4794 rc = whereLoopAddAll(&sWLB); 4795 if( rc ) goto whereBeginError; 4796 4797 #ifdef WHERETRACE_ENABLED 4798 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4799 WhereLoop *p; 4800 int i; 4801 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4802 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4803 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4804 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 4805 whereLoopPrint(p, sWLB.pWC); 4806 } 4807 } 4808 #endif 4809 4810 wherePathSolver(pWInfo, 0); 4811 if( db->mallocFailed ) goto whereBeginError; 4812 if( pWInfo->pOrderBy ){ 4813 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 4814 if( db->mallocFailed ) goto whereBeginError; 4815 } 4816 } 4817 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 4818 pWInfo->revMask = ALLBITS; 4819 } 4820 if( pParse->nErr || NEVER(db->mallocFailed) ){ 4821 goto whereBeginError; 4822 } 4823 #ifdef WHERETRACE_ENABLED 4824 if( sqlite3WhereTrace ){ 4825 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 4826 if( pWInfo->nOBSat>0 ){ 4827 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 4828 } 4829 switch( pWInfo->eDistinct ){ 4830 case WHERE_DISTINCT_UNIQUE: { 4831 sqlite3DebugPrintf(" DISTINCT=unique"); 4832 break; 4833 } 4834 case WHERE_DISTINCT_ORDERED: { 4835 sqlite3DebugPrintf(" DISTINCT=ordered"); 4836 break; 4837 } 4838 case WHERE_DISTINCT_UNORDERED: { 4839 sqlite3DebugPrintf(" DISTINCT=unordered"); 4840 break; 4841 } 4842 } 4843 sqlite3DebugPrintf("\n"); 4844 for(ii=0; ii<pWInfo->nLevel; ii++){ 4845 whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 4846 } 4847 } 4848 #endif 4849 4850 /* Attempt to omit tables from the join that do not affect the result. 4851 ** For a table to not affect the result, the following must be true: 4852 ** 4853 ** 1) The query must not be an aggregate. 4854 ** 2) The table must be the RHS of a LEFT JOIN. 4855 ** 3) Either the query must be DISTINCT, or else the ON or USING clause 4856 ** must contain a constraint that limits the scan of the table to 4857 ** at most a single row. 4858 ** 4) The table must not be referenced by any part of the query apart 4859 ** from its own USING or ON clause. 4860 ** 4861 ** For example, given: 4862 ** 4863 ** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1); 4864 ** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2); 4865 ** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3); 4866 ** 4867 ** then table t2 can be omitted from the following: 4868 ** 4869 ** SELECT v1, v3 FROM t1 4870 ** LEFT JOIN t2 USING (t1.ipk=t2.ipk) 4871 ** LEFT JOIN t3 USING (t1.ipk=t3.ipk) 4872 ** 4873 ** or from: 4874 ** 4875 ** SELECT DISTINCT v1, v3 FROM t1 4876 ** LEFT JOIN t2 4877 ** LEFT JOIN t3 USING (t1.ipk=t3.ipk) 4878 */ 4879 notReady = ~(Bitmask)0; 4880 if( pWInfo->nLevel>=2 4881 && pResultSet!=0 /* guarantees condition (1) above */ 4882 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 4883 ){ 4884 int i; 4885 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet); 4886 if( sWLB.pOrderBy ){ 4887 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy); 4888 } 4889 for(i=pWInfo->nLevel-1; i>=1; i--){ 4890 WhereTerm *pTerm, *pEnd; 4891 struct SrcList_item *pItem; 4892 pLoop = pWInfo->a[i].pWLoop; 4893 pItem = &pWInfo->pTabList->a[pLoop->iTab]; 4894 if( (pItem->fg.jointype & JT_LEFT)==0 ) continue; 4895 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 4896 && (pLoop->wsFlags & WHERE_ONEROW)==0 4897 ){ 4898 continue; 4899 } 4900 if( (tabUsed & pLoop->maskSelf)!=0 ) continue; 4901 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; 4902 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4903 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 4904 if( !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 4905 || pTerm->pExpr->iRightJoinTable!=pItem->iCursor 4906 ){ 4907 break; 4908 } 4909 } 4910 } 4911 if( pTerm<pEnd ) continue; 4912 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 4913 notReady &= ~pLoop->maskSelf; 4914 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4915 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 4916 pTerm->wtFlags |= TERM_CODED; 4917 } 4918 } 4919 if( i!=pWInfo->nLevel-1 ){ 4920 int nByte = (pWInfo->nLevel-1-i) * sizeof(WhereLevel); 4921 memmove(&pWInfo->a[i], &pWInfo->a[i+1], nByte); 4922 } 4923 pWInfo->nLevel--; 4924 nTabList--; 4925 } 4926 } 4927 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 4928 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 4929 4930 /* If the caller is an UPDATE or DELETE statement that is requesting 4931 ** to use a one-pass algorithm, determine if this is appropriate. 4932 ** 4933 ** A one-pass approach can be used if the caller has requested one 4934 ** and either (a) the scan visits at most one row or (b) each 4935 ** of the following are true: 4936 ** 4937 ** * the caller has indicated that a one-pass approach can be used 4938 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and 4939 ** * the table is not a virtual table, and 4940 ** * either the scan does not use the OR optimization or the caller 4941 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified 4942 ** for DELETE). 4943 ** 4944 ** The last qualification is because an UPDATE statement uses 4945 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can 4946 ** use a one-pass approach, and this is not set accurately for scans 4947 ** that use the OR optimization. 4948 */ 4949 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 4950 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 4951 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 4952 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 4953 assert( !(wsFlags & WHERE_VIRTUALTABLE) || IsVirtual(pTabList->a[0].pTab) ); 4954 if( bOnerow || ( 4955 0!=(wctrlFlags & WHERE_ONEPASS_MULTIROW) 4956 && !IsVirtual(pTabList->a[0].pTab) 4957 && (0==(wsFlags & WHERE_MULTI_OR) || (wctrlFlags & WHERE_DUPLICATES_OK)) 4958 )){ 4959 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 4960 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 4961 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 4962 bFordelete = OPFLAG_FORDELETE; 4963 } 4964 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 4965 } 4966 } 4967 } 4968 4969 /* Open all tables in the pTabList and any indices selected for 4970 ** searching those tables. 4971 */ 4972 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 4973 Table *pTab; /* Table to open */ 4974 int iDb; /* Index of database containing table/index */ 4975 struct SrcList_item *pTabItem; 4976 4977 pTabItem = &pTabList->a[pLevel->iFrom]; 4978 pTab = pTabItem->pTab; 4979 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 4980 pLoop = pLevel->pWLoop; 4981 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 4982 /* Do nothing */ 4983 }else 4984 #ifndef SQLITE_OMIT_VIRTUALTABLE 4985 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 4986 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 4987 int iCur = pTabItem->iCursor; 4988 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 4989 }else if( IsVirtual(pTab) ){ 4990 /* noop */ 4991 }else 4992 #endif 4993 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 4994 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 4995 int op = OP_OpenRead; 4996 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4997 op = OP_OpenWrite; 4998 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 4999 }; 5000 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 5001 assert( pTabItem->iCursor==pLevel->iTabCur ); 5002 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 5003 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 5004 if( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol<BMS && HasRowid(pTab) ){ 5005 Bitmask b = pTabItem->colUsed; 5006 int n = 0; 5007 for(; b; b=b>>1, n++){} 5008 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 5009 assert( n<=pTab->nCol ); 5010 } 5011 #ifdef SQLITE_ENABLE_CURSOR_HINTS 5012 if( pLoop->u.btree.pIndex!=0 ){ 5013 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 5014 }else 5015 #endif 5016 { 5017 sqlite3VdbeChangeP5(v, bFordelete); 5018 } 5019 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5020 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 5021 (const u8*)&pTabItem->colUsed, P4_INT64); 5022 #endif 5023 }else{ 5024 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 5025 } 5026 if( pLoop->wsFlags & WHERE_INDEXED ){ 5027 Index *pIx = pLoop->u.btree.pIndex; 5028 int iIndexCur; 5029 int op = OP_OpenRead; 5030 /* iAuxArg is always set to a positive value if ONEPASS is possible */ 5031 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 5032 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 5033 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 5034 ){ 5035 /* This is one term of an OR-optimization using the PRIMARY KEY of a 5036 ** WITHOUT ROWID table. No need for a separate index */ 5037 iIndexCur = pLevel->iTabCur; 5038 op = 0; 5039 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5040 Index *pJ = pTabItem->pTab->pIndex; 5041 iIndexCur = iAuxArg; 5042 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 5043 while( ALWAYS(pJ) && pJ!=pIx ){ 5044 iIndexCur++; 5045 pJ = pJ->pNext; 5046 } 5047 op = OP_OpenWrite; 5048 pWInfo->aiCurOnePass[1] = iIndexCur; 5049 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 5050 iIndexCur = iAuxArg; 5051 op = OP_ReopenIdx; 5052 }else{ 5053 iIndexCur = pParse->nTab++; 5054 } 5055 pLevel->iIdxCur = iIndexCur; 5056 assert( pIx->pSchema==pTab->pSchema ); 5057 assert( iIndexCur>=0 ); 5058 if( op ){ 5059 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 5060 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 5061 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 5062 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 5063 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 5064 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 5065 ){ 5066 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); /* Hint to COMDB2 */ 5067 } 5068 VdbeComment((v, "%s", pIx->zName)); 5069 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5070 { 5071 u64 colUsed = 0; 5072 int ii, jj; 5073 for(ii=0; ii<pIx->nColumn; ii++){ 5074 jj = pIx->aiColumn[ii]; 5075 if( jj<0 ) continue; 5076 if( jj>63 ) jj = 63; 5077 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 5078 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 5079 } 5080 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 5081 (u8*)&colUsed, P4_INT64); 5082 } 5083 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 5084 } 5085 } 5086 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 5087 } 5088 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 5089 if( db->mallocFailed ) goto whereBeginError; 5090 5091 /* Generate the code to do the search. Each iteration of the for 5092 ** loop below generates code for a single nested loop of the VM 5093 ** program. 5094 */ 5095 for(ii=0; ii<nTabList; ii++){ 5096 int addrExplain; 5097 int wsFlags; 5098 pLevel = &pWInfo->a[ii]; 5099 wsFlags = pLevel->pWLoop->wsFlags; 5100 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 5101 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 5102 constructAutomaticIndex(pParse, &pWInfo->sWC, 5103 &pTabList->a[pLevel->iFrom], notReady, pLevel); 5104 if( db->mallocFailed ) goto whereBeginError; 5105 } 5106 #endif 5107 addrExplain = sqlite3WhereExplainOneScan( 5108 pParse, pTabList, pLevel, wctrlFlags 5109 ); 5110 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 5111 notReady = sqlite3WhereCodeOneLoopStart(pParse,v,pWInfo,ii,pLevel,notReady); 5112 pWInfo->iContinue = pLevel->addrCont; 5113 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 5114 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 5115 } 5116 } 5117 5118 /* Done. */ 5119 VdbeModuleComment((v, "Begin WHERE-core")); 5120 return pWInfo; 5121 5122 /* Jump here if malloc fails */ 5123 whereBeginError: 5124 if( pWInfo ){ 5125 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5126 whereInfoFree(db, pWInfo); 5127 } 5128 return 0; 5129 } 5130 5131 /* 5132 ** Part of sqlite3WhereEnd() will rewrite opcodes to reference the 5133 ** index rather than the main table. In SQLITE_DEBUG mode, we want 5134 ** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine 5135 ** does that. 5136 */ 5137 #ifndef SQLITE_DEBUG 5138 # define OpcodeRewriteTrace(D,K,P) /* no-op */ 5139 #else 5140 # define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P) 5141 static void sqlite3WhereOpcodeRewriteTrace( 5142 sqlite3 *db, 5143 int pc, 5144 VdbeOp *pOp 5145 ){ 5146 if( (db->flags & SQLITE_VdbeAddopTrace)==0 ) return; 5147 sqlite3VdbePrintOp(0, pc, pOp); 5148 } 5149 #endif 5150 5151 /* 5152 ** Generate the end of the WHERE loop. See comments on 5153 ** sqlite3WhereBegin() for additional information. 5154 */ 5155 void sqlite3WhereEnd(WhereInfo *pWInfo){ 5156 Parse *pParse = pWInfo->pParse; 5157 Vdbe *v = pParse->pVdbe; 5158 int i; 5159 WhereLevel *pLevel; 5160 WhereLoop *pLoop; 5161 SrcList *pTabList = pWInfo->pTabList; 5162 sqlite3 *db = pParse->db; 5163 5164 /* Generate loop termination code. 5165 */ 5166 VdbeModuleComment((v, "End WHERE-core")); 5167 for(i=pWInfo->nLevel-1; i>=0; i--){ 5168 int addr; 5169 pLevel = &pWInfo->a[i]; 5170 pLoop = pLevel->pWLoop; 5171 if( pLevel->op!=OP_Noop ){ 5172 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5173 int addrSeek = 0; 5174 Index *pIdx; 5175 int n; 5176 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 5177 && i==pWInfo->nLevel-1 /* Ticket [ef9318757b152e3] 2017-10-21 */ 5178 && (pLoop->wsFlags & WHERE_INDEXED)!=0 5179 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 5180 && (n = pLoop->u.btree.nDistinctCol)>0 5181 && pIdx->aiRowLogEst[n]>=36 5182 ){ 5183 int r1 = pParse->nMem+1; 5184 int j, op; 5185 for(j=0; j<n; j++){ 5186 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 5187 } 5188 pParse->nMem += n+1; 5189 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 5190 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 5191 VdbeCoverageIf(v, op==OP_SeekLT); 5192 VdbeCoverageIf(v, op==OP_SeekGT); 5193 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 5194 } 5195 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 5196 /* The common case: Advance to the next row */ 5197 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5198 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 5199 sqlite3VdbeChangeP5(v, pLevel->p5); 5200 VdbeCoverage(v); 5201 VdbeCoverageIf(v, pLevel->op==OP_Next); 5202 VdbeCoverageIf(v, pLevel->op==OP_Prev); 5203 VdbeCoverageIf(v, pLevel->op==OP_VNext); 5204 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5205 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 5206 #endif 5207 }else{ 5208 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5209 } 5210 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 5211 struct InLoop *pIn; 5212 int j; 5213 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 5214 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 5215 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5216 if( pIn->eEndLoopOp!=OP_Noop ){ 5217 if( pIn->nPrefix ){ 5218 assert( pLoop->wsFlags & WHERE_IN_EARLYOUT ); 5219 sqlite3VdbeAddOp4Int(v, OP_IfNoHope, pLevel->iIdxCur, 5220 sqlite3VdbeCurrentAddr(v)+2, 5221 pIn->iBase, pIn->nPrefix); 5222 VdbeCoverage(v); 5223 } 5224 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 5225 VdbeCoverage(v); 5226 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Prev); 5227 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Next); 5228 } 5229 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 5230 } 5231 } 5232 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 5233 if( pLevel->addrSkip ){ 5234 sqlite3VdbeGoto(v, pLevel->addrSkip); 5235 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 5236 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 5237 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 5238 } 5239 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 5240 if( pLevel->addrLikeRep ){ 5241 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 5242 pLevel->addrLikeRep); 5243 VdbeCoverage(v); 5244 } 5245 #endif 5246 if( pLevel->iLeftJoin ){ 5247 int ws = pLoop->wsFlags; 5248 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 5249 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 5250 if( (ws & WHERE_IDX_ONLY)==0 ){ 5251 assert( pLevel->iTabCur==pTabList->a[pLevel->iFrom].iCursor ); 5252 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iTabCur); 5253 } 5254 if( (ws & WHERE_INDEXED) 5255 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCovidx) 5256 ){ 5257 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 5258 } 5259 if( pLevel->op==OP_Return ){ 5260 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 5261 }else{ 5262 sqlite3VdbeGoto(v, pLevel->addrFirst); 5263 } 5264 sqlite3VdbeJumpHere(v, addr); 5265 } 5266 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 5267 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 5268 } 5269 5270 /* The "break" point is here, just past the end of the outer loop. 5271 ** Set it. 5272 */ 5273 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5274 5275 assert( pWInfo->nLevel<=pTabList->nSrc ); 5276 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5277 int k, last; 5278 VdbeOp *pOp; 5279 Index *pIdx = 0; 5280 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; 5281 Table *pTab = pTabItem->pTab; 5282 assert( pTab!=0 ); 5283 pLoop = pLevel->pWLoop; 5284 5285 /* For a co-routine, change all OP_Column references to the table of 5286 ** the co-routine into OP_Copy of result contained in a register. 5287 ** OP_Rowid becomes OP_Null. 5288 */ 5289 if( pTabItem->fg.viaCoroutine ){ 5290 testcase( pParse->db->mallocFailed ); 5291 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 5292 pTabItem->regResult, 0); 5293 continue; 5294 } 5295 5296 #ifdef SQLITE_ENABLE_EARLY_CURSOR_CLOSE 5297 /* Close all of the cursors that were opened by sqlite3WhereBegin. 5298 ** Except, do not close cursors that will be reused by the OR optimization 5299 ** (WHERE_OR_SUBCLAUSE). And do not close the OP_OpenWrite cursors 5300 ** created for the ONEPASS optimization. 5301 */ 5302 if( (pTab->tabFlags & TF_Ephemeral)==0 5303 && pTab->pSelect==0 5304 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5305 ){ 5306 int ws = pLoop->wsFlags; 5307 if( pWInfo->eOnePass==ONEPASS_OFF && (ws & WHERE_IDX_ONLY)==0 ){ 5308 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); 5309 } 5310 if( (ws & WHERE_INDEXED)!=0 5311 && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0 5312 && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1] 5313 ){ 5314 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); 5315 } 5316 } 5317 #endif 5318 5319 /* If this scan uses an index, make VDBE code substitutions to read data 5320 ** from the index instead of from the table where possible. In some cases 5321 ** this optimization prevents the table from ever being read, which can 5322 ** yield a significant performance boost. 5323 ** 5324 ** Calls to the code generator in between sqlite3WhereBegin and 5325 ** sqlite3WhereEnd will have created code that references the table 5326 ** directly. This loop scans all that code looking for opcodes 5327 ** that reference the table and converts them into opcodes that 5328 ** reference the index. 5329 */ 5330 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 5331 pIdx = pLoop->u.btree.pIndex; 5332 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 5333 pIdx = pLevel->u.pCovidx; 5334 } 5335 if( pIdx 5336 && (pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable)) 5337 && !db->mallocFailed 5338 ){ 5339 last = sqlite3VdbeCurrentAddr(v); 5340 k = pLevel->addrBody; 5341 #ifdef SQLITE_DEBUG 5342 if( db->flags & SQLITE_VdbeAddopTrace ){ 5343 printf("TRANSLATE opcodes in range %d..%d\n", k, last-1); 5344 } 5345 #endif 5346 pOp = sqlite3VdbeGetOp(v, k); 5347 for(; k<last; k++, pOp++){ 5348 if( pOp->p1!=pLevel->iTabCur ) continue; 5349 if( pOp->opcode==OP_Column 5350 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC 5351 || pOp->opcode==OP_Offset 5352 #endif 5353 ){ 5354 int x = pOp->p2; 5355 assert( pIdx->pTable==pTab ); 5356 if( !HasRowid(pTab) ){ 5357 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 5358 x = pPk->aiColumn[x]; 5359 assert( x>=0 ); 5360 } 5361 x = sqlite3ColumnOfIndex(pIdx, x); 5362 if( x>=0 ){ 5363 pOp->p2 = x; 5364 pOp->p1 = pLevel->iIdxCur; 5365 OpcodeRewriteTrace(db, k, pOp); 5366 } 5367 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 5368 || pWInfo->eOnePass ); 5369 }else if( pOp->opcode==OP_Rowid ){ 5370 pOp->p1 = pLevel->iIdxCur; 5371 pOp->opcode = OP_IdxRowid; 5372 OpcodeRewriteTrace(db, k, pOp); 5373 }else if( pOp->opcode==OP_IfNullRow ){ 5374 pOp->p1 = pLevel->iIdxCur; 5375 OpcodeRewriteTrace(db, k, pOp); 5376 } 5377 } 5378 #ifdef SQLITE_DEBUG 5379 if( db->flags & SQLITE_VdbeAddopTrace ) printf("TRANSLATE complete\n"); 5380 #endif 5381 } 5382 } 5383 5384 /* Final cleanup 5385 */ 5386 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5387 whereInfoFree(db, pWInfo); 5388 return; 5389 } 5390