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 ** Initialize a WHERE clause scanner object. Return a pointer to the 315 ** first match. Return NULL if there are no matches. 316 ** 317 ** The scanner will be searching the WHERE clause pWC. It will look 318 ** for terms of the form "X <op> <expr>" where X is column iColumn of table 319 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx 320 ** must be one of the indexes of table iCur. 321 ** 322 ** The <op> must be one of the operators described by opMask. 323 ** 324 ** If the search is for X and the WHERE clause contains terms of the 325 ** form X=Y then this routine might also return terms of the form 326 ** "Y <op> <expr>". The number of levels of transitivity is limited, 327 ** but is enough to handle most commonly occurring SQL statements. 328 ** 329 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with 330 ** index pIdx. 331 */ 332 static WhereTerm *whereScanInit( 333 WhereScan *pScan, /* The WhereScan object being initialized */ 334 WhereClause *pWC, /* The WHERE clause to be scanned */ 335 int iCur, /* Cursor to scan for */ 336 int iColumn, /* Column to scan for */ 337 u32 opMask, /* Operator(s) to scan for */ 338 Index *pIdx /* Must be compatible with this index */ 339 ){ 340 pScan->pOrigWC = pWC; 341 pScan->pWC = pWC; 342 pScan->pIdxExpr = 0; 343 pScan->idxaff = 0; 344 pScan->zCollName = 0; 345 if( pIdx ){ 346 int j = iColumn; 347 iColumn = pIdx->aiColumn[j]; 348 if( iColumn==XN_EXPR ){ 349 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr; 350 pScan->zCollName = pIdx->azColl[j]; 351 }else if( iColumn==pIdx->pTable->iPKey ){ 352 iColumn = XN_ROWID; 353 }else if( iColumn>=0 ){ 354 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; 355 pScan->zCollName = pIdx->azColl[j]; 356 } 357 }else if( iColumn==XN_EXPR ){ 358 return 0; 359 } 360 pScan->opMask = opMask; 361 pScan->k = 0; 362 pScan->aiCur[0] = iCur; 363 pScan->aiColumn[0] = iColumn; 364 pScan->nEquiv = 1; 365 pScan->iEquiv = 1; 366 return whereScanNext(pScan); 367 } 368 369 /* 370 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 371 ** where X is a reference to the iColumn of table iCur or of index pIdx 372 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by 373 ** the op parameter. Return a pointer to the term. Return 0 if not found. 374 ** 375 ** If pIdx!=0 then it must be one of the indexes of table iCur. 376 ** Search for terms matching the iColumn-th column of pIdx 377 ** rather than the iColumn-th column of table iCur. 378 ** 379 ** The term returned might by Y=<expr> if there is another constraint in 380 ** the WHERE clause that specifies that X=Y. Any such constraints will be 381 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 382 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11 383 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10 384 ** other equivalent values. Hence a search for X will return <expr> if X=A1 385 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>. 386 ** 387 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 388 ** then try for the one with no dependencies on <expr> - in other words where 389 ** <expr> is a constant expression of some kind. Only return entries of 390 ** the form "X <op> Y" where Y is a column in another table if no terms of 391 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 392 ** exist, try to return a term that does not use WO_EQUIV. 393 */ 394 WhereTerm *sqlite3WhereFindTerm( 395 WhereClause *pWC, /* The WHERE clause to be searched */ 396 int iCur, /* Cursor number of LHS */ 397 int iColumn, /* Column number of LHS */ 398 Bitmask notReady, /* RHS must not overlap with this mask */ 399 u32 op, /* Mask of WO_xx values describing operator */ 400 Index *pIdx /* Must be compatible with this index, if not NULL */ 401 ){ 402 WhereTerm *pResult = 0; 403 WhereTerm *p; 404 WhereScan scan; 405 406 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); 407 op &= WO_EQ|WO_IS; 408 while( p ){ 409 if( (p->prereqRight & notReady)==0 ){ 410 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){ 411 testcase( p->eOperator & WO_IS ); 412 return p; 413 } 414 if( pResult==0 ) pResult = p; 415 } 416 p = whereScanNext(&scan); 417 } 418 return pResult; 419 } 420 421 /* 422 ** This function searches pList for an entry that matches the iCol-th column 423 ** of index pIdx. 424 ** 425 ** If such an expression is found, its index in pList->a[] is returned. If 426 ** no expression is found, -1 is returned. 427 */ 428 static int findIndexCol( 429 Parse *pParse, /* Parse context */ 430 ExprList *pList, /* Expression list to search */ 431 int iBase, /* Cursor for table associated with pIdx */ 432 Index *pIdx, /* Index to match column of */ 433 int iCol /* Column of index to match */ 434 ){ 435 int i; 436 const char *zColl = pIdx->azColl[iCol]; 437 438 for(i=0; i<pList->nExpr; i++){ 439 Expr *p = sqlite3ExprSkipCollate(pList->a[i].pExpr); 440 if( p->op==TK_COLUMN 441 && p->iColumn==pIdx->aiColumn[iCol] 442 && p->iTable==iBase 443 ){ 444 CollSeq *pColl = sqlite3ExprNNCollSeq(pParse, pList->a[i].pExpr); 445 if( 0==sqlite3StrICmp(pColl->zName, zColl) ){ 446 return i; 447 } 448 } 449 } 450 451 return -1; 452 } 453 454 /* 455 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL 456 */ 457 static int indexColumnNotNull(Index *pIdx, int iCol){ 458 int j; 459 assert( pIdx!=0 ); 460 assert( iCol>=0 && iCol<pIdx->nColumn ); 461 j = pIdx->aiColumn[iCol]; 462 if( j>=0 ){ 463 return pIdx->pTable->aCol[j].notNull; 464 }else if( j==(-1) ){ 465 return 1; 466 }else{ 467 assert( j==(-2) ); 468 return 0; /* Assume an indexed expression can always yield a NULL */ 469 470 } 471 } 472 473 /* 474 ** Return true if the DISTINCT expression-list passed as the third argument 475 ** is redundant. 476 ** 477 ** A DISTINCT list is redundant if any subset of the columns in the 478 ** DISTINCT list are collectively unique and individually non-null. 479 */ 480 static int isDistinctRedundant( 481 Parse *pParse, /* Parsing context */ 482 SrcList *pTabList, /* The FROM clause */ 483 WhereClause *pWC, /* The WHERE clause */ 484 ExprList *pDistinct /* The result set that needs to be DISTINCT */ 485 ){ 486 Table *pTab; 487 Index *pIdx; 488 int i; 489 int iBase; 490 491 /* If there is more than one table or sub-select in the FROM clause of 492 ** this query, then it will not be possible to show that the DISTINCT 493 ** clause is redundant. */ 494 if( pTabList->nSrc!=1 ) return 0; 495 iBase = pTabList->a[0].iCursor; 496 pTab = pTabList->a[0].pTab; 497 498 /* If any of the expressions is an IPK column on table iBase, then return 499 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 500 ** current SELECT is a correlated sub-query. 501 */ 502 for(i=0; i<pDistinct->nExpr; i++){ 503 Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr); 504 if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1; 505 } 506 507 /* Loop through all indices on the table, checking each to see if it makes 508 ** the DISTINCT qualifier redundant. It does so if: 509 ** 510 ** 1. The index is itself UNIQUE, and 511 ** 512 ** 2. All of the columns in the index are either part of the pDistinct 513 ** list, or else the WHERE clause contains a term of the form "col=X", 514 ** where X is a constant value. The collation sequences of the 515 ** comparison and select-list expressions must match those of the index. 516 ** 517 ** 3. All of those index columns for which the WHERE clause does not 518 ** contain a "col=X" term are subject to a NOT NULL constraint. 519 */ 520 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 521 if( !IsUniqueIndex(pIdx) ) continue; 522 for(i=0; i<pIdx->nKeyCol; i++){ 523 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){ 524 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break; 525 if( indexColumnNotNull(pIdx, i)==0 ) break; 526 } 527 } 528 if( i==pIdx->nKeyCol ){ 529 /* This index implies that the DISTINCT qualifier is redundant. */ 530 return 1; 531 } 532 } 533 534 return 0; 535 } 536 537 538 /* 539 ** Estimate the logarithm of the input value to base 2. 540 */ 541 static LogEst estLog(LogEst N){ 542 return N<=10 ? 0 : sqlite3LogEst(N) - 33; 543 } 544 545 /* 546 ** Convert OP_Column opcodes to OP_Copy in previously generated code. 547 ** 548 ** This routine runs over generated VDBE code and translates OP_Column 549 ** opcodes into OP_Copy when the table is being accessed via co-routine 550 ** instead of via table lookup. 551 ** 552 ** If the bIncrRowid parameter is 0, then any OP_Rowid instructions on 553 ** cursor iTabCur are transformed into OP_Null. Or, if bIncrRowid is non-zero, 554 ** then each OP_Rowid is transformed into an instruction to increment the 555 ** value stored in its output register. 556 */ 557 static void translateColumnToCopy( 558 Parse *pParse, /* Parsing context */ 559 int iStart, /* Translate from this opcode to the end */ 560 int iTabCur, /* OP_Column/OP_Rowid references to this table */ 561 int iRegister, /* The first column is in this register */ 562 int bIncrRowid /* If non-zero, transform OP_rowid to OP_AddImm(1) */ 563 ){ 564 Vdbe *v = pParse->pVdbe; 565 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart); 566 int iEnd = sqlite3VdbeCurrentAddr(v); 567 if( pParse->db->mallocFailed ) return; 568 for(; iStart<iEnd; iStart++, pOp++){ 569 if( pOp->p1!=iTabCur ) continue; 570 if( pOp->opcode==OP_Column ){ 571 pOp->opcode = OP_Copy; 572 pOp->p1 = pOp->p2 + iRegister; 573 pOp->p2 = pOp->p3; 574 pOp->p3 = 0; 575 }else if( pOp->opcode==OP_Rowid ){ 576 if( bIncrRowid ){ 577 /* Increment the value stored in the P2 operand of the OP_Rowid. */ 578 pOp->opcode = OP_AddImm; 579 pOp->p1 = pOp->p2; 580 pOp->p2 = 1; 581 }else{ 582 pOp->opcode = OP_Null; 583 pOp->p1 = 0; 584 pOp->p3 = 0; 585 } 586 } 587 } 588 } 589 590 /* 591 ** Two routines for printing the content of an sqlite3_index_info 592 ** structure. Used for testing and debugging only. If neither 593 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 594 ** are no-ops. 595 */ 596 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) 597 static void TRACE_IDX_INPUTS(sqlite3_index_info *p){ 598 int i; 599 if( !sqlite3WhereTrace ) return; 600 for(i=0; i<p->nConstraint; i++){ 601 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", 602 i, 603 p->aConstraint[i].iColumn, 604 p->aConstraint[i].iTermOffset, 605 p->aConstraint[i].op, 606 p->aConstraint[i].usable); 607 } 608 for(i=0; i<p->nOrderBy; i++){ 609 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 610 i, 611 p->aOrderBy[i].iColumn, 612 p->aOrderBy[i].desc); 613 } 614 } 615 static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){ 616 int i; 617 if( !sqlite3WhereTrace ) return; 618 for(i=0; i<p->nConstraint; i++){ 619 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 620 i, 621 p->aConstraintUsage[i].argvIndex, 622 p->aConstraintUsage[i].omit); 623 } 624 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 625 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 626 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 627 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 628 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); 629 } 630 #else 631 #define TRACE_IDX_INPUTS(A) 632 #define TRACE_IDX_OUTPUTS(A) 633 #endif 634 635 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 636 /* 637 ** Return TRUE if the WHERE clause term pTerm is of a form where it 638 ** could be used with an index to access pSrc, assuming an appropriate 639 ** index existed. 640 */ 641 static int termCanDriveIndex( 642 WhereTerm *pTerm, /* WHERE clause term to check */ 643 struct SrcList_item *pSrc, /* Table we are trying to access */ 644 Bitmask notReady /* Tables in outer loops of the join */ 645 ){ 646 char aff; 647 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 648 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0; 649 if( (pSrc->fg.jointype & JT_LEFT) 650 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 651 && (pTerm->eOperator & WO_IS) 652 ){ 653 /* Cannot use an IS term from the WHERE clause as an index driver for 654 ** the RHS of a LEFT JOIN. Such a term can only be used if it is from 655 ** the ON clause. */ 656 return 0; 657 } 658 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 659 if( pTerm->u.leftColumn<0 ) return 0; 660 aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity; 661 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 662 testcase( pTerm->pExpr->op==TK_IS ); 663 return 1; 664 } 665 #endif 666 667 668 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 669 /* 670 ** Generate code to construct the Index object for an automatic index 671 ** and to set up the WhereLevel object pLevel so that the code generator 672 ** makes use of the automatic index. 673 */ 674 static void constructAutomaticIndex( 675 Parse *pParse, /* The parsing context */ 676 WhereClause *pWC, /* The WHERE clause */ 677 struct SrcList_item *pSrc, /* The FROM clause term to get the next index */ 678 Bitmask notReady, /* Mask of cursors that are not available */ 679 WhereLevel *pLevel /* Write new index here */ 680 ){ 681 int nKeyCol; /* Number of columns in the constructed index */ 682 WhereTerm *pTerm; /* A single term of the WHERE clause */ 683 WhereTerm *pWCEnd; /* End of pWC->a[] */ 684 Index *pIdx; /* Object describing the transient index */ 685 Vdbe *v; /* Prepared statement under construction */ 686 int addrInit; /* Address of the initialization bypass jump */ 687 Table *pTable; /* The table being indexed */ 688 int addrTop; /* Top of the index fill loop */ 689 int regRecord; /* Register holding an index record */ 690 int n; /* Column counter */ 691 int i; /* Loop counter */ 692 int mxBitCol; /* Maximum column in pSrc->colUsed */ 693 CollSeq *pColl; /* Collating sequence to on a column */ 694 WhereLoop *pLoop; /* The Loop object */ 695 char *zNotUsed; /* Extra space on the end of pIdx */ 696 Bitmask idxCols; /* Bitmap of columns used for indexing */ 697 Bitmask extraCols; /* Bitmap of additional columns */ 698 u8 sentWarning = 0; /* True if a warnning has been issued */ 699 Expr *pPartial = 0; /* Partial Index Expression */ 700 int iContinue = 0; /* Jump here to skip excluded rows */ 701 struct SrcList_item *pTabItem; /* FROM clause term being indexed */ 702 int addrCounter = 0; /* Address where integer counter is initialized */ 703 int regBase; /* Array of registers where record is assembled */ 704 705 /* Generate code to skip over the creation and initialization of the 706 ** transient index on 2nd and subsequent iterations of the loop. */ 707 v = pParse->pVdbe; 708 assert( v!=0 ); 709 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 710 711 /* Count the number of columns that will be added to the index 712 ** and used to match WHERE clause constraints */ 713 nKeyCol = 0; 714 pTable = pSrc->pTab; 715 pWCEnd = &pWC->a[pWC->nTerm]; 716 pLoop = pLevel->pWLoop; 717 idxCols = 0; 718 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 719 Expr *pExpr = pTerm->pExpr; 720 assert( !ExprHasProperty(pExpr, EP_FromJoin) /* prereq always non-zero */ 721 || pExpr->iRightJoinTable!=pSrc->iCursor /* for the right-hand */ 722 || pLoop->prereq!=0 ); /* table of a LEFT JOIN */ 723 if( pLoop->prereq==0 724 && (pTerm->wtFlags & TERM_VIRTUAL)==0 725 && !ExprHasProperty(pExpr, EP_FromJoin) 726 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) ){ 727 pPartial = sqlite3ExprAnd(pParse->db, pPartial, 728 sqlite3ExprDup(pParse->db, pExpr, 0)); 729 } 730 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 731 int iCol = pTerm->u.leftColumn; 732 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 733 testcase( iCol==BMS ); 734 testcase( iCol==BMS-1 ); 735 if( !sentWarning ){ 736 sqlite3_log(SQLITE_WARNING_AUTOINDEX, 737 "automatic index on %s(%s)", pTable->zName, 738 pTable->aCol[iCol].zName); 739 sentWarning = 1; 740 } 741 if( (idxCols & cMask)==0 ){ 742 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ 743 goto end_auto_index_create; 744 } 745 pLoop->aLTerm[nKeyCol++] = pTerm; 746 idxCols |= cMask; 747 } 748 } 749 } 750 assert( nKeyCol>0 ); 751 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; 752 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED 753 | WHERE_AUTO_INDEX; 754 755 /* Count the number of additional columns needed to create a 756 ** covering index. A "covering index" is an index that contains all 757 ** columns that are needed by the query. With a covering index, the 758 ** original table never needs to be accessed. Automatic indices must 759 ** be a covering index because the index will not be updated if the 760 ** original table changes and the index and table cannot both be used 761 ** if they go out of sync. 762 */ 763 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); 764 mxBitCol = MIN(BMS-1,pTable->nCol); 765 testcase( pTable->nCol==BMS-1 ); 766 testcase( pTable->nCol==BMS-2 ); 767 for(i=0; i<mxBitCol; i++){ 768 if( extraCols & MASKBIT(i) ) nKeyCol++; 769 } 770 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 771 nKeyCol += pTable->nCol - BMS + 1; 772 } 773 774 /* Construct the Index object to describe this index */ 775 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); 776 if( pIdx==0 ) goto end_auto_index_create; 777 pLoop->u.btree.pIndex = pIdx; 778 pIdx->zName = "auto-index"; 779 pIdx->pTable = pTable; 780 n = 0; 781 idxCols = 0; 782 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 783 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 784 int iCol = pTerm->u.leftColumn; 785 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 786 testcase( iCol==BMS-1 ); 787 testcase( iCol==BMS ); 788 if( (idxCols & cMask)==0 ){ 789 Expr *pX = pTerm->pExpr; 790 idxCols |= cMask; 791 pIdx->aiColumn[n] = pTerm->u.leftColumn; 792 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); 793 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY; 794 n++; 795 } 796 } 797 } 798 assert( (u32)n==pLoop->u.btree.nEq ); 799 800 /* Add additional columns needed to make the automatic index into 801 ** a covering index */ 802 for(i=0; i<mxBitCol; i++){ 803 if( extraCols & MASKBIT(i) ){ 804 pIdx->aiColumn[n] = i; 805 pIdx->azColl[n] = sqlite3StrBINARY; 806 n++; 807 } 808 } 809 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 810 for(i=BMS-1; i<pTable->nCol; i++){ 811 pIdx->aiColumn[n] = i; 812 pIdx->azColl[n] = sqlite3StrBINARY; 813 n++; 814 } 815 } 816 assert( n==nKeyCol ); 817 pIdx->aiColumn[n] = XN_ROWID; 818 pIdx->azColl[n] = sqlite3StrBINARY; 819 820 /* Create the automatic index */ 821 assert( pLevel->iIdxCur>=0 ); 822 pLevel->iIdxCur = pParse->nTab++; 823 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); 824 sqlite3VdbeSetP4KeyInfo(pParse, pIdx); 825 VdbeComment((v, "for %s", pTable->zName)); 826 827 /* Fill the automatic index with content */ 828 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom]; 829 if( pTabItem->fg.viaCoroutine ){ 830 int regYield = pTabItem->regReturn; 831 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0); 832 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); 833 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 834 VdbeCoverage(v); 835 VdbeComment((v, "next row of %s", pTabItem->pTab->zName)); 836 }else{ 837 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); 838 } 839 if( pPartial ){ 840 iContinue = sqlite3VdbeMakeLabel(v); 841 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); 842 pLoop->wsFlags |= WHERE_PARTIALIDX; 843 } 844 regRecord = sqlite3GetTempReg(pParse); 845 regBase = sqlite3GenerateIndexKey( 846 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0 847 ); 848 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 849 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 850 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); 851 if( pTabItem->fg.viaCoroutine ){ 852 sqlite3VdbeChangeP2(v, addrCounter, regBase+n); 853 testcase( pParse->db->mallocFailed ); 854 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur, 855 pTabItem->regResult, 1); 856 sqlite3VdbeGoto(v, addrTop); 857 }else{ 858 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); 859 } 860 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 861 sqlite3VdbeJumpHere(v, addrTop); 862 sqlite3ReleaseTempReg(pParse, regRecord); 863 864 /* Jump here when skipping the initialization */ 865 sqlite3VdbeJumpHere(v, addrInit); 866 867 end_auto_index_create: 868 sqlite3ExprDelete(pParse->db, pPartial); 869 } 870 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 871 872 #ifndef SQLITE_OMIT_VIRTUALTABLE 873 /* 874 ** Allocate and populate an sqlite3_index_info structure. It is the 875 ** responsibility of the caller to eventually release the structure 876 ** by passing the pointer returned by this function to sqlite3_free(). 877 */ 878 static sqlite3_index_info *allocateIndexInfo( 879 Parse *pParse, /* The parsing context */ 880 WhereClause *pWC, /* The WHERE clause being analyzed */ 881 Bitmask mUnusable, /* Ignore terms with these prereqs */ 882 struct SrcList_item *pSrc, /* The FROM clause term that is the vtab */ 883 ExprList *pOrderBy, /* The ORDER BY clause */ 884 u16 *pmNoOmit /* Mask of terms not to omit */ 885 ){ 886 int i, j; 887 int nTerm; 888 struct sqlite3_index_constraint *pIdxCons; 889 struct sqlite3_index_orderby *pIdxOrderBy; 890 struct sqlite3_index_constraint_usage *pUsage; 891 struct HiddenIndexInfo *pHidden; 892 WhereTerm *pTerm; 893 int nOrderBy; 894 sqlite3_index_info *pIdxInfo; 895 u16 mNoOmit = 0; 896 897 /* Count the number of possible WHERE clause constraints referring 898 ** to this virtual table */ 899 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 900 if( pTerm->leftCursor != pSrc->iCursor ) continue; 901 if( pTerm->prereqRight & mUnusable ) continue; 902 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 903 testcase( pTerm->eOperator & WO_IN ); 904 testcase( pTerm->eOperator & WO_ISNULL ); 905 testcase( pTerm->eOperator & WO_IS ); 906 testcase( pTerm->eOperator & WO_ALL ); 907 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 908 if( pTerm->wtFlags & TERM_VNULL ) continue; 909 assert( pTerm->u.leftColumn>=(-1) ); 910 nTerm++; 911 } 912 913 /* If the ORDER BY clause contains only columns in the current 914 ** virtual table then allocate space for the aOrderBy part of 915 ** the sqlite3_index_info structure. 916 */ 917 nOrderBy = 0; 918 if( pOrderBy ){ 919 int n = pOrderBy->nExpr; 920 for(i=0; i<n; i++){ 921 Expr *pExpr = pOrderBy->a[i].pExpr; 922 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; 923 } 924 if( i==n){ 925 nOrderBy = n; 926 } 927 } 928 929 /* Allocate the sqlite3_index_info structure 930 */ 931 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 932 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 933 + sizeof(*pIdxOrderBy)*nOrderBy + sizeof(*pHidden) ); 934 if( pIdxInfo==0 ){ 935 sqlite3ErrorMsg(pParse, "out of memory"); 936 return 0; 937 } 938 939 /* Initialize the structure. The sqlite3_index_info structure contains 940 ** many fields that are declared "const" to prevent xBestIndex from 941 ** changing them. We have to do some funky casting in order to 942 ** initialize those fields. 943 */ 944 pHidden = (struct HiddenIndexInfo*)&pIdxInfo[1]; 945 pIdxCons = (struct sqlite3_index_constraint*)&pHidden[1]; 946 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 947 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 948 *(int*)&pIdxInfo->nConstraint = nTerm; 949 *(int*)&pIdxInfo->nOrderBy = nOrderBy; 950 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons; 951 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy; 952 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage = 953 pUsage; 954 955 pHidden->pWC = pWC; 956 pHidden->pParse = pParse; 957 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 958 u16 op; 959 if( pTerm->leftCursor != pSrc->iCursor ) continue; 960 if( pTerm->prereqRight & mUnusable ) continue; 961 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 962 testcase( pTerm->eOperator & WO_IN ); 963 testcase( pTerm->eOperator & WO_IS ); 964 testcase( pTerm->eOperator & WO_ISNULL ); 965 testcase( pTerm->eOperator & WO_ALL ); 966 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 967 if( pTerm->wtFlags & TERM_VNULL ) continue; 968 if( (pSrc->fg.jointype & JT_LEFT)!=0 969 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 970 && (pTerm->eOperator & (WO_IS|WO_ISNULL)) 971 ){ 972 /* An "IS" term in the WHERE clause where the virtual table is the rhs 973 ** of a LEFT JOIN. Do not pass this term to the virtual table 974 ** implementation, as this can lead to incorrect results from SQL such 975 ** as: 976 ** 977 ** "LEFT JOIN vtab WHERE vtab.col IS NULL" */ 978 testcase( pTerm->eOperator & WO_ISNULL ); 979 testcase( pTerm->eOperator & WO_IS ); 980 continue; 981 } 982 assert( pTerm->u.leftColumn>=(-1) ); 983 pIdxCons[j].iColumn = pTerm->u.leftColumn; 984 pIdxCons[j].iTermOffset = i; 985 op = pTerm->eOperator & WO_ALL; 986 if( op==WO_IN ) op = WO_EQ; 987 if( op==WO_AUX ){ 988 pIdxCons[j].op = pTerm->eMatchOp; 989 }else if( op & (WO_ISNULL|WO_IS) ){ 990 if( op==WO_ISNULL ){ 991 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_ISNULL; 992 }else{ 993 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_IS; 994 } 995 }else{ 996 pIdxCons[j].op = (u8)op; 997 /* The direct assignment in the previous line is possible only because 998 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 999 ** following asserts verify this fact. */ 1000 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 1001 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 1002 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 1003 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 1004 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 1005 assert( pTerm->eOperator&(WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_AUX) ); 1006 1007 if( op & (WO_LT|WO_LE|WO_GT|WO_GE) 1008 && sqlite3ExprIsVector(pTerm->pExpr->pRight) 1009 ){ 1010 if( i<16 ) mNoOmit |= (1 << i); 1011 if( op==WO_LT ) pIdxCons[j].op = WO_LE; 1012 if( op==WO_GT ) pIdxCons[j].op = WO_GE; 1013 } 1014 } 1015 1016 j++; 1017 } 1018 for(i=0; i<nOrderBy; i++){ 1019 Expr *pExpr = pOrderBy->a[i].pExpr; 1020 pIdxOrderBy[i].iColumn = pExpr->iColumn; 1021 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder; 1022 } 1023 1024 *pmNoOmit = mNoOmit; 1025 return pIdxInfo; 1026 } 1027 1028 /* 1029 ** The table object reference passed as the second argument to this function 1030 ** must represent a virtual table. This function invokes the xBestIndex() 1031 ** method of the virtual table with the sqlite3_index_info object that 1032 ** comes in as the 3rd argument to this function. 1033 ** 1034 ** If an error occurs, pParse is populated with an error message and an 1035 ** appropriate error code is returned. A return of SQLITE_CONSTRAINT from 1036 ** xBestIndex is not considered an error. SQLITE_CONSTRAINT indicates that 1037 ** the current configuration of "unusable" flags in sqlite3_index_info can 1038 ** not result in a valid plan. 1039 ** 1040 ** Whether or not an error is returned, it is the responsibility of the 1041 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 1042 ** that this is required. 1043 */ 1044 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 1045 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 1046 int rc; 1047 1048 TRACE_IDX_INPUTS(p); 1049 rc = pVtab->pModule->xBestIndex(pVtab, p); 1050 TRACE_IDX_OUTPUTS(p); 1051 1052 if( rc!=SQLITE_OK && rc!=SQLITE_CONSTRAINT ){ 1053 if( rc==SQLITE_NOMEM ){ 1054 sqlite3OomFault(pParse->db); 1055 }else if( !pVtab->zErrMsg ){ 1056 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 1057 }else{ 1058 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 1059 } 1060 } 1061 sqlite3_free(pVtab->zErrMsg); 1062 pVtab->zErrMsg = 0; 1063 return rc; 1064 } 1065 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 1066 1067 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1068 /* 1069 ** Estimate the location of a particular key among all keys in an 1070 ** index. Store the results in aStat as follows: 1071 ** 1072 ** aStat[0] Est. number of rows less than pRec 1073 ** aStat[1] Est. number of rows equal to pRec 1074 ** 1075 ** Return the index of the sample that is the smallest sample that 1076 ** is greater than or equal to pRec. Note that this index is not an index 1077 ** into the aSample[] array - it is an index into a virtual set of samples 1078 ** based on the contents of aSample[] and the number of fields in record 1079 ** pRec. 1080 */ 1081 static int whereKeyStats( 1082 Parse *pParse, /* Database connection */ 1083 Index *pIdx, /* Index to consider domain of */ 1084 UnpackedRecord *pRec, /* Vector of values to consider */ 1085 int roundUp, /* Round up if true. Round down if false */ 1086 tRowcnt *aStat /* OUT: stats written here */ 1087 ){ 1088 IndexSample *aSample = pIdx->aSample; 1089 int iCol; /* Index of required stats in anEq[] etc. */ 1090 int i; /* Index of first sample >= pRec */ 1091 int iSample; /* Smallest sample larger than or equal to pRec */ 1092 int iMin = 0; /* Smallest sample not yet tested */ 1093 int iTest; /* Next sample to test */ 1094 int res; /* Result of comparison operation */ 1095 int nField; /* Number of fields in pRec */ 1096 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 1097 1098 #ifndef SQLITE_DEBUG 1099 UNUSED_PARAMETER( pParse ); 1100 #endif 1101 assert( pRec!=0 ); 1102 assert( pIdx->nSample>0 ); 1103 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1104 1105 /* Do a binary search to find the first sample greater than or equal 1106 ** to pRec. If pRec contains a single field, the set of samples to search 1107 ** is simply the aSample[] array. If the samples in aSample[] contain more 1108 ** than one fields, all fields following the first are ignored. 1109 ** 1110 ** If pRec contains N fields, where N is more than one, then as well as the 1111 ** samples in aSample[] (truncated to N fields), the search also has to 1112 ** consider prefixes of those samples. For example, if the set of samples 1113 ** in aSample is: 1114 ** 1115 ** aSample[0] = (a, 5) 1116 ** aSample[1] = (a, 10) 1117 ** aSample[2] = (b, 5) 1118 ** aSample[3] = (c, 100) 1119 ** aSample[4] = (c, 105) 1120 ** 1121 ** Then the search space should ideally be the samples above and the 1122 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1123 ** the code actually searches this set: 1124 ** 1125 ** 0: (a) 1126 ** 1: (a, 5) 1127 ** 2: (a, 10) 1128 ** 3: (a, 10) 1129 ** 4: (b) 1130 ** 5: (b, 5) 1131 ** 6: (c) 1132 ** 7: (c, 100) 1133 ** 8: (c, 105) 1134 ** 9: (c, 105) 1135 ** 1136 ** For each sample in the aSample[] array, N samples are present in the 1137 ** effective sample array. In the above, samples 0 and 1 are based on 1138 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1139 ** 1140 ** Often, sample i of each block of N effective samples has (i+1) fields. 1141 ** Except, each sample may be extended to ensure that it is greater than or 1142 ** equal to the previous sample in the array. For example, in the above, 1143 ** sample 2 is the first sample of a block of N samples, so at first it 1144 ** appears that it should be 1 field in size. However, that would make it 1145 ** smaller than sample 1, so the binary search would not work. As a result, 1146 ** it is extended to two fields. The duplicates that this creates do not 1147 ** cause any problems. 1148 */ 1149 nField = pRec->nField; 1150 iCol = 0; 1151 iSample = pIdx->nSample * nField; 1152 do{ 1153 int iSamp; /* Index in aSample[] of test sample */ 1154 int n; /* Number of fields in test sample */ 1155 1156 iTest = (iMin+iSample)/2; 1157 iSamp = iTest / nField; 1158 if( iSamp>0 ){ 1159 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1160 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1161 ** fields that is greater than the previous effective sample. */ 1162 for(n=(iTest % nField) + 1; n<nField; n++){ 1163 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1164 } 1165 }else{ 1166 n = iTest + 1; 1167 } 1168 1169 pRec->nField = n; 1170 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1171 if( res<0 ){ 1172 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1173 iMin = iTest+1; 1174 }else if( res==0 && n<nField ){ 1175 iLower = aSample[iSamp].anLt[n-1]; 1176 iMin = iTest+1; 1177 res = -1; 1178 }else{ 1179 iSample = iTest; 1180 iCol = n-1; 1181 } 1182 }while( res && iMin<iSample ); 1183 i = iSample / nField; 1184 1185 #ifdef SQLITE_DEBUG 1186 /* The following assert statements check that the binary search code 1187 ** above found the right answer. This block serves no purpose other 1188 ** than to invoke the asserts. */ 1189 if( pParse->db->mallocFailed==0 ){ 1190 if( res==0 ){ 1191 /* If (res==0) is true, then pRec must be equal to sample i. */ 1192 assert( i<pIdx->nSample ); 1193 assert( iCol==nField-1 ); 1194 pRec->nField = nField; 1195 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1196 || pParse->db->mallocFailed 1197 ); 1198 }else{ 1199 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1200 ** all samples in the aSample[] array, pRec must be smaller than the 1201 ** (iCol+1) field prefix of sample i. */ 1202 assert( i<=pIdx->nSample && i>=0 ); 1203 pRec->nField = iCol+1; 1204 assert( i==pIdx->nSample 1205 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1206 || pParse->db->mallocFailed ); 1207 1208 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1209 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1210 ** be greater than or equal to the (iCol) field prefix of sample i. 1211 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1212 if( iCol>0 ){ 1213 pRec->nField = iCol; 1214 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1215 || pParse->db->mallocFailed ); 1216 } 1217 if( i>0 ){ 1218 pRec->nField = nField; 1219 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1220 || pParse->db->mallocFailed ); 1221 } 1222 } 1223 } 1224 #endif /* ifdef SQLITE_DEBUG */ 1225 1226 if( res==0 ){ 1227 /* Record pRec is equal to sample i */ 1228 assert( iCol==nField-1 ); 1229 aStat[0] = aSample[i].anLt[iCol]; 1230 aStat[1] = aSample[i].anEq[iCol]; 1231 }else{ 1232 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1233 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1234 ** is larger than all samples in the array. */ 1235 tRowcnt iUpper, iGap; 1236 if( i>=pIdx->nSample ){ 1237 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1238 }else{ 1239 iUpper = aSample[i].anLt[iCol]; 1240 } 1241 1242 if( iLower>=iUpper ){ 1243 iGap = 0; 1244 }else{ 1245 iGap = iUpper - iLower; 1246 } 1247 if( roundUp ){ 1248 iGap = (iGap*2)/3; 1249 }else{ 1250 iGap = iGap/3; 1251 } 1252 aStat[0] = iLower + iGap; 1253 aStat[1] = pIdx->aAvgEq[nField-1]; 1254 } 1255 1256 /* Restore the pRec->nField value before returning. */ 1257 pRec->nField = nField; 1258 return i; 1259 } 1260 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1261 1262 /* 1263 ** If it is not NULL, pTerm is a term that provides an upper or lower 1264 ** bound on a range scan. Without considering pTerm, it is estimated 1265 ** that the scan will visit nNew rows. This function returns the number 1266 ** estimated to be visited after taking pTerm into account. 1267 ** 1268 ** If the user explicitly specified a likelihood() value for this term, 1269 ** then the return value is the likelihood multiplied by the number of 1270 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1271 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1272 */ 1273 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1274 LogEst nRet = nNew; 1275 if( pTerm ){ 1276 if( pTerm->truthProb<=0 ){ 1277 nRet += pTerm->truthProb; 1278 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1279 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1280 } 1281 } 1282 return nRet; 1283 } 1284 1285 1286 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1287 /* 1288 ** Return the affinity for a single column of an index. 1289 */ 1290 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1291 assert( iCol>=0 && iCol<pIdx->nColumn ); 1292 if( !pIdx->zColAff ){ 1293 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1294 } 1295 return pIdx->zColAff[iCol]; 1296 } 1297 #endif 1298 1299 1300 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1301 /* 1302 ** This function is called to estimate the number of rows visited by a 1303 ** range-scan on a skip-scan index. For example: 1304 ** 1305 ** CREATE INDEX i1 ON t1(a, b, c); 1306 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1307 ** 1308 ** Value pLoop->nOut is currently set to the estimated number of rows 1309 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1310 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1311 ** on the stat4 data for the index. this scan will be peformed multiple 1312 ** times (once for each (a,b) combination that matches a=?) is dealt with 1313 ** by the caller. 1314 ** 1315 ** It does this by scanning through all stat4 samples, comparing values 1316 ** extracted from pLower and pUpper with the corresponding column in each 1317 ** sample. If L and U are the number of samples found to be less than or 1318 ** equal to the values extracted from pLower and pUpper respectively, and 1319 ** N is the total number of samples, the pLoop->nOut value is adjusted 1320 ** as follows: 1321 ** 1322 ** nOut = nOut * ( min(U - L, 1) / N ) 1323 ** 1324 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1325 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1326 ** U is set to N. 1327 ** 1328 ** Normally, this function sets *pbDone to 1 before returning. However, 1329 ** if no value can be extracted from either pLower or pUpper (and so the 1330 ** estimate of the number of rows delivered remains unchanged), *pbDone 1331 ** is left as is. 1332 ** 1333 ** If an error occurs, an SQLite error code is returned. Otherwise, 1334 ** SQLITE_OK. 1335 */ 1336 static int whereRangeSkipScanEst( 1337 Parse *pParse, /* Parsing & code generating context */ 1338 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1339 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1340 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1341 int *pbDone /* Set to true if at least one expr. value extracted */ 1342 ){ 1343 Index *p = pLoop->u.btree.pIndex; 1344 int nEq = pLoop->u.btree.nEq; 1345 sqlite3 *db = pParse->db; 1346 int nLower = -1; 1347 int nUpper = p->nSample+1; 1348 int rc = SQLITE_OK; 1349 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1350 CollSeq *pColl; 1351 1352 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1353 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1354 sqlite3_value *pVal = 0; /* Value extracted from record */ 1355 1356 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1357 if( pLower ){ 1358 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1359 nLower = 0; 1360 } 1361 if( pUpper && rc==SQLITE_OK ){ 1362 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1363 nUpper = p2 ? 0 : p->nSample; 1364 } 1365 1366 if( p1 || p2 ){ 1367 int i; 1368 int nDiff; 1369 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1370 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1371 if( rc==SQLITE_OK && p1 ){ 1372 int res = sqlite3MemCompare(p1, pVal, pColl); 1373 if( res>=0 ) nLower++; 1374 } 1375 if( rc==SQLITE_OK && p2 ){ 1376 int res = sqlite3MemCompare(p2, pVal, pColl); 1377 if( res>=0 ) nUpper++; 1378 } 1379 } 1380 nDiff = (nUpper - nLower); 1381 if( nDiff<=0 ) nDiff = 1; 1382 1383 /* If there is both an upper and lower bound specified, and the 1384 ** comparisons indicate that they are close together, use the fallback 1385 ** method (assume that the scan visits 1/64 of the rows) for estimating 1386 ** the number of rows visited. Otherwise, estimate the number of rows 1387 ** using the method described in the header comment for this function. */ 1388 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1389 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1390 pLoop->nOut -= nAdjust; 1391 *pbDone = 1; 1392 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1393 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1394 } 1395 1396 }else{ 1397 assert( *pbDone==0 ); 1398 } 1399 1400 sqlite3ValueFree(p1); 1401 sqlite3ValueFree(p2); 1402 sqlite3ValueFree(pVal); 1403 1404 return rc; 1405 } 1406 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1407 1408 /* 1409 ** This function is used to estimate the number of rows that will be visited 1410 ** by scanning an index for a range of values. The range may have an upper 1411 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1412 ** and lower bounds are represented by pLower and pUpper respectively. For 1413 ** example, assuming that index p is on t1(a): 1414 ** 1415 ** ... FROM t1 WHERE a > ? AND a < ? ... 1416 ** |_____| |_____| 1417 ** | | 1418 ** pLower pUpper 1419 ** 1420 ** If either of the upper or lower bound is not present, then NULL is passed in 1421 ** place of the corresponding WhereTerm. 1422 ** 1423 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1424 ** column subject to the range constraint. Or, equivalently, the number of 1425 ** equality constraints optimized by the proposed index scan. For example, 1426 ** assuming index p is on t1(a, b), and the SQL query is: 1427 ** 1428 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1429 ** 1430 ** then nEq is set to 1 (as the range restricted column, b, is the second 1431 ** left-most column of the index). Or, if the query is: 1432 ** 1433 ** ... FROM t1 WHERE a > ? AND a < ? ... 1434 ** 1435 ** then nEq is set to 0. 1436 ** 1437 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1438 ** number of rows that the index scan is expected to visit without 1439 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1440 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1441 ** to account for the range constraints pLower and pUpper. 1442 ** 1443 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1444 ** used, a single range inequality reduces the search space by a factor of 4. 1445 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1446 ** rows visited by a factor of 64. 1447 */ 1448 static int whereRangeScanEst( 1449 Parse *pParse, /* Parsing & code generating context */ 1450 WhereLoopBuilder *pBuilder, 1451 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1452 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1453 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1454 ){ 1455 int rc = SQLITE_OK; 1456 int nOut = pLoop->nOut; 1457 LogEst nNew; 1458 1459 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1460 Index *p = pLoop->u.btree.pIndex; 1461 int nEq = pLoop->u.btree.nEq; 1462 1463 if( p->nSample>0 && nEq<p->nSampleCol 1464 && OptimizationEnabled(pParse->db, SQLITE_Stat34) 1465 ){ 1466 if( nEq==pBuilder->nRecValid ){ 1467 UnpackedRecord *pRec = pBuilder->pRec; 1468 tRowcnt a[2]; 1469 int nBtm = pLoop->u.btree.nBtm; 1470 int nTop = pLoop->u.btree.nTop; 1471 1472 /* Variable iLower will be set to the estimate of the number of rows in 1473 ** the index that are less than the lower bound of the range query. The 1474 ** lower bound being the concatenation of $P and $L, where $P is the 1475 ** key-prefix formed by the nEq values matched against the nEq left-most 1476 ** columns of the index, and $L is the value in pLower. 1477 ** 1478 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1479 ** is not a simple variable or literal value), the lower bound of the 1480 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1481 ** if $L is available, whereKeyStats() is called for both ($P) and 1482 ** ($P:$L) and the larger of the two returned values is used. 1483 ** 1484 ** Similarly, iUpper is to be set to the estimate of the number of rows 1485 ** less than the upper bound of the range query. Where the upper bound 1486 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1487 ** of iUpper are requested of whereKeyStats() and the smaller used. 1488 ** 1489 ** The number of rows between the two bounds is then just iUpper-iLower. 1490 */ 1491 tRowcnt iLower; /* Rows less than the lower bound */ 1492 tRowcnt iUpper; /* Rows less than the upper bound */ 1493 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1494 int iUprIdx = -1; /* aSample[] for the upper bound */ 1495 1496 if( pRec ){ 1497 testcase( pRec->nField!=pBuilder->nRecValid ); 1498 pRec->nField = pBuilder->nRecValid; 1499 } 1500 /* Determine iLower and iUpper using ($P) only. */ 1501 if( nEq==0 ){ 1502 iLower = 0; 1503 iUpper = p->nRowEst0; 1504 }else{ 1505 /* Note: this call could be optimized away - since the same values must 1506 ** have been requested when testing key $P in whereEqualScanEst(). */ 1507 whereKeyStats(pParse, p, pRec, 0, a); 1508 iLower = a[0]; 1509 iUpper = a[0] + a[1]; 1510 } 1511 1512 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1513 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1514 assert( p->aSortOrder!=0 ); 1515 if( p->aSortOrder[nEq] ){ 1516 /* The roles of pLower and pUpper are swapped for a DESC index */ 1517 SWAP(WhereTerm*, pLower, pUpper); 1518 SWAP(int, nBtm, nTop); 1519 } 1520 1521 /* If possible, improve on the iLower estimate using ($P:$L). */ 1522 if( pLower ){ 1523 int n; /* Values extracted from pExpr */ 1524 Expr *pExpr = pLower->pExpr->pRight; 1525 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1526 if( rc==SQLITE_OK && n ){ 1527 tRowcnt iNew; 1528 u16 mask = WO_GT|WO_LE; 1529 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1530 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1531 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1532 if( iNew>iLower ) iLower = iNew; 1533 nOut--; 1534 pLower = 0; 1535 } 1536 } 1537 1538 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1539 if( pUpper ){ 1540 int n; /* Values extracted from pExpr */ 1541 Expr *pExpr = pUpper->pExpr->pRight; 1542 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1543 if( rc==SQLITE_OK && n ){ 1544 tRowcnt iNew; 1545 u16 mask = WO_GT|WO_LE; 1546 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1547 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1548 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1549 if( iNew<iUpper ) iUpper = iNew; 1550 nOut--; 1551 pUpper = 0; 1552 } 1553 } 1554 1555 pBuilder->pRec = pRec; 1556 if( rc==SQLITE_OK ){ 1557 if( iUpper>iLower ){ 1558 nNew = sqlite3LogEst(iUpper - iLower); 1559 /* TUNING: If both iUpper and iLower are derived from the same 1560 ** sample, then assume they are 4x more selective. This brings 1561 ** the estimated selectivity more in line with what it would be 1562 ** if estimated without the use of STAT3/4 tables. */ 1563 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1564 }else{ 1565 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1566 } 1567 if( nNew<nOut ){ 1568 nOut = nNew; 1569 } 1570 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1571 (u32)iLower, (u32)iUpper, nOut)); 1572 } 1573 }else{ 1574 int bDone = 0; 1575 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1576 if( bDone ) return rc; 1577 } 1578 } 1579 #else 1580 UNUSED_PARAMETER(pParse); 1581 UNUSED_PARAMETER(pBuilder); 1582 assert( pLower || pUpper ); 1583 #endif 1584 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1585 nNew = whereRangeAdjust(pLower, nOut); 1586 nNew = whereRangeAdjust(pUpper, nNew); 1587 1588 /* TUNING: If there is both an upper and lower limit and neither limit 1589 ** has an application-defined likelihood(), assume the range is 1590 ** reduced by an additional 75%. This means that, by default, an open-ended 1591 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1592 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1593 ** match 1/64 of the index. */ 1594 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1595 nNew -= 20; 1596 } 1597 1598 nOut -= (pLower!=0) + (pUpper!=0); 1599 if( nNew<10 ) nNew = 10; 1600 if( nNew<nOut ) nOut = nNew; 1601 #if defined(WHERETRACE_ENABLED) 1602 if( pLoop->nOut>nOut ){ 1603 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1604 pLoop->nOut, nOut)); 1605 } 1606 #endif 1607 pLoop->nOut = (LogEst)nOut; 1608 return rc; 1609 } 1610 1611 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1612 /* 1613 ** Estimate the number of rows that will be returned based on 1614 ** an equality constraint x=VALUE and where that VALUE occurs in 1615 ** the histogram data. This only works when x is the left-most 1616 ** column of an index and sqlite_stat3 histogram data is available 1617 ** for that index. When pExpr==NULL that means the constraint is 1618 ** "x IS NULL" instead of "x=VALUE". 1619 ** 1620 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1621 ** If unable to make an estimate, leave *pnRow unchanged and return 1622 ** non-zero. 1623 ** 1624 ** This routine can fail if it is unable to load a collating sequence 1625 ** required for string comparison, or if unable to allocate memory 1626 ** for a UTF conversion required for comparison. The error is stored 1627 ** in the pParse structure. 1628 */ 1629 static int whereEqualScanEst( 1630 Parse *pParse, /* Parsing & code generating context */ 1631 WhereLoopBuilder *pBuilder, 1632 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1633 tRowcnt *pnRow /* Write the revised row estimate here */ 1634 ){ 1635 Index *p = pBuilder->pNew->u.btree.pIndex; 1636 int nEq = pBuilder->pNew->u.btree.nEq; 1637 UnpackedRecord *pRec = pBuilder->pRec; 1638 int rc; /* Subfunction return code */ 1639 tRowcnt a[2]; /* Statistics */ 1640 int bOk; 1641 1642 assert( nEq>=1 ); 1643 assert( nEq<=p->nColumn ); 1644 assert( p->aSample!=0 ); 1645 assert( p->nSample>0 ); 1646 assert( pBuilder->nRecValid<nEq ); 1647 1648 /* If values are not available for all fields of the index to the left 1649 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1650 if( pBuilder->nRecValid<(nEq-1) ){ 1651 return SQLITE_NOTFOUND; 1652 } 1653 1654 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1655 ** below would return the same value. */ 1656 if( nEq>=p->nColumn ){ 1657 *pnRow = 1; 1658 return SQLITE_OK; 1659 } 1660 1661 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1662 pBuilder->pRec = pRec; 1663 if( rc!=SQLITE_OK ) return rc; 1664 if( bOk==0 ) return SQLITE_NOTFOUND; 1665 pBuilder->nRecValid = nEq; 1666 1667 whereKeyStats(pParse, p, pRec, 0, a); 1668 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1669 p->zName, nEq-1, (int)a[1])); 1670 *pnRow = a[1]; 1671 1672 return rc; 1673 } 1674 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1675 1676 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 1677 /* 1678 ** Estimate the number of rows that will be returned based on 1679 ** an IN constraint where the right-hand side of the IN operator 1680 ** is a list of values. Example: 1681 ** 1682 ** WHERE x IN (1,2,3,4) 1683 ** 1684 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1685 ** If unable to make an estimate, leave *pnRow unchanged and return 1686 ** non-zero. 1687 ** 1688 ** This routine can fail if it is unable to load a collating sequence 1689 ** required for string comparison, or if unable to allocate memory 1690 ** for a UTF conversion required for comparison. The error is stored 1691 ** in the pParse structure. 1692 */ 1693 static int whereInScanEst( 1694 Parse *pParse, /* Parsing & code generating context */ 1695 WhereLoopBuilder *pBuilder, 1696 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1697 tRowcnt *pnRow /* Write the revised row estimate here */ 1698 ){ 1699 Index *p = pBuilder->pNew->u.btree.pIndex; 1700 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1701 int nRecValid = pBuilder->nRecValid; 1702 int rc = SQLITE_OK; /* Subfunction return code */ 1703 tRowcnt nEst; /* Number of rows for a single term */ 1704 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1705 int i; /* Loop counter */ 1706 1707 assert( p->aSample!=0 ); 1708 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1709 nEst = nRow0; 1710 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1711 nRowEst += nEst; 1712 pBuilder->nRecValid = nRecValid; 1713 } 1714 1715 if( rc==SQLITE_OK ){ 1716 if( nRowEst > nRow0 ) nRowEst = nRow0; 1717 *pnRow = nRowEst; 1718 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1719 } 1720 assert( pBuilder->nRecValid==nRecValid ); 1721 return rc; 1722 } 1723 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ 1724 1725 1726 #ifdef WHERETRACE_ENABLED 1727 /* 1728 ** Print the content of a WhereTerm object 1729 */ 1730 static void whereTermPrint(WhereTerm *pTerm, int iTerm){ 1731 if( pTerm==0 ){ 1732 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1733 }else{ 1734 char zType[4]; 1735 char zLeft[50]; 1736 memcpy(zType, "...", 4); 1737 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1738 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1739 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1740 if( pTerm->eOperator & WO_SINGLE ){ 1741 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 1742 pTerm->leftCursor, pTerm->u.leftColumn); 1743 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 1744 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%lld", 1745 pTerm->u.pOrInfo->indexable); 1746 }else{ 1747 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 1748 } 1749 sqlite3DebugPrintf( 1750 "TERM-%-3d %p %s %-12s prob=%-3d op=0x%03x wtFlags=0x%04x", 1751 iTerm, pTerm, zType, zLeft, pTerm->truthProb, 1752 pTerm->eOperator, pTerm->wtFlags); 1753 if( pTerm->iField ){ 1754 sqlite3DebugPrintf(" iField=%d\n", pTerm->iField); 1755 }else{ 1756 sqlite3DebugPrintf("\n"); 1757 } 1758 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 1759 } 1760 } 1761 #endif 1762 1763 #ifdef WHERETRACE_ENABLED 1764 /* 1765 ** Show the complete content of a WhereClause 1766 */ 1767 void sqlite3WhereClausePrint(WhereClause *pWC){ 1768 int i; 1769 for(i=0; i<pWC->nTerm; i++){ 1770 whereTermPrint(&pWC->a[i], i); 1771 } 1772 } 1773 #endif 1774 1775 #ifdef WHERETRACE_ENABLED 1776 /* 1777 ** Print a WhereLoop object for debugging purposes 1778 */ 1779 static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){ 1780 WhereInfo *pWInfo = pWC->pWInfo; 1781 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 1782 struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab; 1783 Table *pTab = pItem->pTab; 1784 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 1785 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 1786 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 1787 sqlite3DebugPrintf(" %12s", 1788 pItem->zAlias ? pItem->zAlias : pTab->zName); 1789 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 1790 const char *zName; 1791 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 1792 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 1793 int i = sqlite3Strlen30(zName) - 1; 1794 while( zName[i]!='_' ) i--; 1795 zName += i; 1796 } 1797 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 1798 }else{ 1799 sqlite3DebugPrintf("%20s",""); 1800 } 1801 }else{ 1802 char *z; 1803 if( p->u.vtab.idxStr ){ 1804 z = sqlite3_mprintf("(%d,\"%s\",%x)", 1805 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 1806 }else{ 1807 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 1808 } 1809 sqlite3DebugPrintf(" %-19s", z); 1810 sqlite3_free(z); 1811 } 1812 if( p->wsFlags & WHERE_SKIPSCAN ){ 1813 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 1814 }else{ 1815 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); 1816 } 1817 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 1818 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 1819 int i; 1820 for(i=0; i<p->nLTerm; i++){ 1821 whereTermPrint(p->aLTerm[i], i); 1822 } 1823 } 1824 } 1825 #endif 1826 1827 /* 1828 ** Convert bulk memory into a valid WhereLoop that can be passed 1829 ** to whereLoopClear harmlessly. 1830 */ 1831 static void whereLoopInit(WhereLoop *p){ 1832 p->aLTerm = p->aLTermSpace; 1833 p->nLTerm = 0; 1834 p->nLSlot = ArraySize(p->aLTermSpace); 1835 p->wsFlags = 0; 1836 } 1837 1838 /* 1839 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 1840 */ 1841 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 1842 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 1843 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 1844 sqlite3_free(p->u.vtab.idxStr); 1845 p->u.vtab.needFree = 0; 1846 p->u.vtab.idxStr = 0; 1847 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 1848 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 1849 sqlite3DbFreeNN(db, p->u.btree.pIndex); 1850 p->u.btree.pIndex = 0; 1851 } 1852 } 1853 } 1854 1855 /* 1856 ** Deallocate internal memory used by a WhereLoop object 1857 */ 1858 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 1859 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1860 whereLoopClearUnion(db, p); 1861 whereLoopInit(p); 1862 } 1863 1864 /* 1865 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 1866 */ 1867 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 1868 WhereTerm **paNew; 1869 if( p->nLSlot>=n ) return SQLITE_OK; 1870 n = (n+7)&~7; 1871 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 1872 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 1873 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 1874 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1875 p->aLTerm = paNew; 1876 p->nLSlot = n; 1877 return SQLITE_OK; 1878 } 1879 1880 /* 1881 ** Transfer content from the second pLoop into the first. 1882 */ 1883 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 1884 whereLoopClearUnion(db, pTo); 1885 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 1886 memset(&pTo->u, 0, sizeof(pTo->u)); 1887 return SQLITE_NOMEM_BKPT; 1888 } 1889 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 1890 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 1891 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 1892 pFrom->u.vtab.needFree = 0; 1893 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 1894 pFrom->u.btree.pIndex = 0; 1895 } 1896 return SQLITE_OK; 1897 } 1898 1899 /* 1900 ** Delete a WhereLoop object 1901 */ 1902 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 1903 whereLoopClear(db, p); 1904 sqlite3DbFreeNN(db, p); 1905 } 1906 1907 /* 1908 ** Free a WhereInfo structure 1909 */ 1910 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 1911 int i; 1912 assert( pWInfo!=0 ); 1913 for(i=0; i<pWInfo->nLevel; i++){ 1914 WhereLevel *pLevel = &pWInfo->a[i]; 1915 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){ 1916 sqlite3DbFree(db, pLevel->u.in.aInLoop); 1917 } 1918 } 1919 sqlite3WhereClauseClear(&pWInfo->sWC); 1920 while( pWInfo->pLoops ){ 1921 WhereLoop *p = pWInfo->pLoops; 1922 pWInfo->pLoops = p->pNextLoop; 1923 whereLoopDelete(db, p); 1924 } 1925 sqlite3DbFreeNN(db, pWInfo); 1926 } 1927 1928 /* 1929 ** Return TRUE if all of the following are true: 1930 ** 1931 ** (1) X has the same or lower cost that Y 1932 ** (2) X uses fewer WHERE clause terms than Y 1933 ** (3) Every WHERE clause term used by X is also used by Y 1934 ** (4) X skips at least as many columns as Y 1935 ** (5) If X is a covering index, than Y is too 1936 ** 1937 ** Conditions (2) and (3) mean that X is a "proper subset" of Y. 1938 ** If X is a proper subset of Y then Y is a better choice and ought 1939 ** to have a lower cost. This routine returns TRUE when that cost 1940 ** relationship is inverted and needs to be adjusted. Constraint (4) 1941 ** was added because if X uses skip-scan less than Y it still might 1942 ** deserve a lower cost even if it is a proper subset of Y. Constraint (5) 1943 ** was added because a covering index probably deserves to have a lower cost 1944 ** than a non-covering index even if it is a proper subset. 1945 */ 1946 static int whereLoopCheaperProperSubset( 1947 const WhereLoop *pX, /* First WhereLoop to compare */ 1948 const WhereLoop *pY /* Compare against this WhereLoop */ 1949 ){ 1950 int i, j; 1951 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 1952 return 0; /* X is not a subset of Y */ 1953 } 1954 if( pY->nSkip > pX->nSkip ) return 0; 1955 if( pX->rRun >= pY->rRun ){ 1956 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ 1957 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ 1958 } 1959 for(i=pX->nLTerm-1; i>=0; i--){ 1960 if( pX->aLTerm[i]==0 ) continue; 1961 for(j=pY->nLTerm-1; j>=0; j--){ 1962 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 1963 } 1964 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 1965 } 1966 if( (pX->wsFlags&WHERE_IDX_ONLY)!=0 1967 && (pY->wsFlags&WHERE_IDX_ONLY)==0 ){ 1968 return 0; /* Constraint (5) */ 1969 } 1970 return 1; /* All conditions meet */ 1971 } 1972 1973 /* 1974 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so 1975 ** that: 1976 ** 1977 ** (1) pTemplate costs less than any other WhereLoops that are a proper 1978 ** subset of pTemplate 1979 ** 1980 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 1981 ** is a proper subset. 1982 ** 1983 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 1984 ** WHERE clause terms than Y and that every WHERE clause term used by X is 1985 ** also used by Y. 1986 */ 1987 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 1988 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 1989 for(; p; p=p->pNextLoop){ 1990 if( p->iTab!=pTemplate->iTab ) continue; 1991 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 1992 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 1993 /* Adjust pTemplate cost downward so that it is cheaper than its 1994 ** subset p. */ 1995 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 1996 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); 1997 pTemplate->rRun = p->rRun; 1998 pTemplate->nOut = p->nOut - 1; 1999 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 2000 /* Adjust pTemplate cost upward so that it is costlier than p since 2001 ** pTemplate is a proper subset of p */ 2002 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2003 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); 2004 pTemplate->rRun = p->rRun; 2005 pTemplate->nOut = p->nOut + 1; 2006 } 2007 } 2008 } 2009 2010 /* 2011 ** Search the list of WhereLoops in *ppPrev looking for one that can be 2012 ** replaced by pTemplate. 2013 ** 2014 ** Return NULL if pTemplate does not belong on the WhereLoop list. 2015 ** In other words if pTemplate ought to be dropped from further consideration. 2016 ** 2017 ** If pX is a WhereLoop that pTemplate can replace, then return the 2018 ** link that points to pX. 2019 ** 2020 ** If pTemplate cannot replace any existing element of the list but needs 2021 ** to be added to the list as a new entry, then return a pointer to the 2022 ** tail of the list. 2023 */ 2024 static WhereLoop **whereLoopFindLesser( 2025 WhereLoop **ppPrev, 2026 const WhereLoop *pTemplate 2027 ){ 2028 WhereLoop *p; 2029 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 2030 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 2031 /* If either the iTab or iSortIdx values for two WhereLoop are different 2032 ** then those WhereLoops need to be considered separately. Neither is 2033 ** a candidate to replace the other. */ 2034 continue; 2035 } 2036 /* In the current implementation, the rSetup value is either zero 2037 ** or the cost of building an automatic index (NlogN) and the NlogN 2038 ** is the same for compatible WhereLoops. */ 2039 assert( p->rSetup==0 || pTemplate->rSetup==0 2040 || p->rSetup==pTemplate->rSetup ); 2041 2042 /* whereLoopAddBtree() always generates and inserts the automatic index 2043 ** case first. Hence compatible candidate WhereLoops never have a larger 2044 ** rSetup. Call this SETUP-INVARIANT */ 2045 assert( p->rSetup>=pTemplate->rSetup ); 2046 2047 /* Any loop using an appliation-defined index (or PRIMARY KEY or 2048 ** UNIQUE constraint) with one or more == constraints is better 2049 ** than an automatic index. Unless it is a skip-scan. */ 2050 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 2051 && (pTemplate->nSkip)==0 2052 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 2053 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 2054 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2055 ){ 2056 break; 2057 } 2058 2059 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2060 ** discarded. WhereLoop p is better if: 2061 ** (1) p has no more dependencies than pTemplate, and 2062 ** (2) p has an equal or lower cost than pTemplate 2063 */ 2064 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2065 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2066 && p->rRun<=pTemplate->rRun /* (2b) */ 2067 && p->nOut<=pTemplate->nOut /* (2c) */ 2068 ){ 2069 return 0; /* Discard pTemplate */ 2070 } 2071 2072 /* If pTemplate is always better than p, then cause p to be overwritten 2073 ** with pTemplate. pTemplate is better than p if: 2074 ** (1) pTemplate has no more dependences than p, and 2075 ** (2) pTemplate has an equal or lower cost than p. 2076 */ 2077 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2078 && p->rRun>=pTemplate->rRun /* (2a) */ 2079 && p->nOut>=pTemplate->nOut /* (2b) */ 2080 ){ 2081 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2082 break; /* Cause p to be overwritten by pTemplate */ 2083 } 2084 } 2085 return ppPrev; 2086 } 2087 2088 /* 2089 ** Insert or replace a WhereLoop entry using the template supplied. 2090 ** 2091 ** An existing WhereLoop entry might be overwritten if the new template 2092 ** is better and has fewer dependencies. Or the template will be ignored 2093 ** and no insert will occur if an existing WhereLoop is faster and has 2094 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2095 ** added based on the template. 2096 ** 2097 ** If pBuilder->pOrSet is not NULL then we care about only the 2098 ** prerequisites and rRun and nOut costs of the N best loops. That 2099 ** information is gathered in the pBuilder->pOrSet object. This special 2100 ** processing mode is used only for OR clause processing. 2101 ** 2102 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2103 ** still might overwrite similar loops with the new template if the 2104 ** new template is better. Loops may be overwritten if the following 2105 ** conditions are met: 2106 ** 2107 ** (1) They have the same iTab. 2108 ** (2) They have the same iSortIdx. 2109 ** (3) The template has same or fewer dependencies than the current loop 2110 ** (4) The template has the same or lower cost than the current loop 2111 */ 2112 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2113 WhereLoop **ppPrev, *p; 2114 WhereInfo *pWInfo = pBuilder->pWInfo; 2115 sqlite3 *db = pWInfo->pParse->db; 2116 int rc; 2117 2118 /* Stop the search once we hit the query planner search limit */ 2119 if( pBuilder->iPlanLimit==0 ){ 2120 WHERETRACE(0xffffffff,("=== query planner search limit reached ===\n")); 2121 if( pBuilder->pOrSet ) pBuilder->pOrSet->n = 0; 2122 return SQLITE_DONE; 2123 } 2124 pBuilder->iPlanLimit--; 2125 2126 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2127 ** and prereqs. 2128 */ 2129 if( pBuilder->pOrSet!=0 ){ 2130 if( pTemplate->nLTerm ){ 2131 #if WHERETRACE_ENABLED 2132 u16 n = pBuilder->pOrSet->n; 2133 int x = 2134 #endif 2135 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2136 pTemplate->nOut); 2137 #if WHERETRACE_ENABLED /* 0x8 */ 2138 if( sqlite3WhereTrace & 0x8 ){ 2139 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2140 whereLoopPrint(pTemplate, pBuilder->pWC); 2141 } 2142 #endif 2143 } 2144 return SQLITE_OK; 2145 } 2146 2147 /* Look for an existing WhereLoop to replace with pTemplate 2148 */ 2149 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2150 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2151 2152 if( ppPrev==0 ){ 2153 /* There already exists a WhereLoop on the list that is better 2154 ** than pTemplate, so just ignore pTemplate */ 2155 #if WHERETRACE_ENABLED /* 0x8 */ 2156 if( sqlite3WhereTrace & 0x8 ){ 2157 sqlite3DebugPrintf(" skip: "); 2158 whereLoopPrint(pTemplate, pBuilder->pWC); 2159 } 2160 #endif 2161 return SQLITE_OK; 2162 }else{ 2163 p = *ppPrev; 2164 } 2165 2166 /* If we reach this point it means that either p[] should be overwritten 2167 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2168 ** WhereLoop and insert it. 2169 */ 2170 #if WHERETRACE_ENABLED /* 0x8 */ 2171 if( sqlite3WhereTrace & 0x8 ){ 2172 if( p!=0 ){ 2173 sqlite3DebugPrintf("replace: "); 2174 whereLoopPrint(p, pBuilder->pWC); 2175 sqlite3DebugPrintf(" with: "); 2176 }else{ 2177 sqlite3DebugPrintf(" add: "); 2178 } 2179 whereLoopPrint(pTemplate, pBuilder->pWC); 2180 } 2181 #endif 2182 if( p==0 ){ 2183 /* Allocate a new WhereLoop to add to the end of the list */ 2184 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2185 if( p==0 ) return SQLITE_NOMEM_BKPT; 2186 whereLoopInit(p); 2187 p->pNextLoop = 0; 2188 }else{ 2189 /* We will be overwriting WhereLoop p[]. But before we do, first 2190 ** go through the rest of the list and delete any other entries besides 2191 ** p[] that are also supplated by pTemplate */ 2192 WhereLoop **ppTail = &p->pNextLoop; 2193 WhereLoop *pToDel; 2194 while( *ppTail ){ 2195 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2196 if( ppTail==0 ) break; 2197 pToDel = *ppTail; 2198 if( pToDel==0 ) break; 2199 *ppTail = pToDel->pNextLoop; 2200 #if WHERETRACE_ENABLED /* 0x8 */ 2201 if( sqlite3WhereTrace & 0x8 ){ 2202 sqlite3DebugPrintf(" delete: "); 2203 whereLoopPrint(pToDel, pBuilder->pWC); 2204 } 2205 #endif 2206 whereLoopDelete(db, pToDel); 2207 } 2208 } 2209 rc = whereLoopXfer(db, p, pTemplate); 2210 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2211 Index *pIndex = p->u.btree.pIndex; 2212 if( pIndex && pIndex->tnum==0 ){ 2213 p->u.btree.pIndex = 0; 2214 } 2215 } 2216 return rc; 2217 } 2218 2219 /* 2220 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2221 ** WHERE clause that reference the loop but which are not used by an 2222 ** index. 2223 * 2224 ** For every WHERE clause term that is not used by the index 2225 ** and which has a truth probability assigned by one of the likelihood(), 2226 ** likely(), or unlikely() SQL functions, reduce the estimated number 2227 ** of output rows by the probability specified. 2228 ** 2229 ** TUNING: For every WHERE clause term that is not used by the index 2230 ** and which does not have an assigned truth probability, heuristics 2231 ** described below are used to try to estimate the truth probability. 2232 ** TODO --> Perhaps this is something that could be improved by better 2233 ** table statistics. 2234 ** 2235 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2236 ** value corresponds to -1 in LogEst notation, so this means decrement 2237 ** the WhereLoop.nOut field for every such WHERE clause term. 2238 ** 2239 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2240 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2241 ** final output row estimate is no greater than 1/4 of the total number 2242 ** of rows in the table. In other words, assume that x==EXPR will filter 2243 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2244 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2245 ** on the "x" column and so in that case only cap the output row estimate 2246 ** at 1/2 instead of 1/4. 2247 */ 2248 static void whereLoopOutputAdjust( 2249 WhereClause *pWC, /* The WHERE clause */ 2250 WhereLoop *pLoop, /* The loop to adjust downward */ 2251 LogEst nRow /* Number of rows in the entire table */ 2252 ){ 2253 WhereTerm *pTerm, *pX; 2254 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2255 int i, j, k; 2256 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2257 2258 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2259 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ 2260 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; 2261 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2262 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2263 for(j=pLoop->nLTerm-1; j>=0; j--){ 2264 pX = pLoop->aLTerm[j]; 2265 if( pX==0 ) continue; 2266 if( pX==pTerm ) break; 2267 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2268 } 2269 if( j<0 ){ 2270 if( pTerm->truthProb<=0 ){ 2271 /* If a truth probability is specified using the likelihood() hints, 2272 ** then use the probability provided by the application. */ 2273 pLoop->nOut += pTerm->truthProb; 2274 }else{ 2275 /* In the absence of explicit truth probabilities, use heuristics to 2276 ** guess a reasonable truth probability. */ 2277 pLoop->nOut--; 2278 if( pTerm->eOperator&(WO_EQ|WO_IS) ){ 2279 Expr *pRight = pTerm->pExpr->pRight; 2280 testcase( pTerm->pExpr->op==TK_IS ); 2281 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2282 k = 10; 2283 }else{ 2284 k = 20; 2285 } 2286 if( iReduce<k ) iReduce = k; 2287 } 2288 } 2289 } 2290 } 2291 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce; 2292 } 2293 2294 /* 2295 ** Term pTerm is a vector range comparison operation. The first comparison 2296 ** in the vector can be optimized using column nEq of the index. This 2297 ** function returns the total number of vector elements that can be used 2298 ** as part of the range comparison. 2299 ** 2300 ** For example, if the query is: 2301 ** 2302 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2303 ** 2304 ** and the index: 2305 ** 2306 ** CREATE INDEX ... ON (a, b, c, d, e) 2307 ** 2308 ** then this function would be invoked with nEq=1. The value returned in 2309 ** this case is 3. 2310 */ 2311 static int whereRangeVectorLen( 2312 Parse *pParse, /* Parsing context */ 2313 int iCur, /* Cursor open on pIdx */ 2314 Index *pIdx, /* The index to be used for a inequality constraint */ 2315 int nEq, /* Number of prior equality constraints on same index */ 2316 WhereTerm *pTerm /* The vector inequality constraint */ 2317 ){ 2318 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2319 int i; 2320 2321 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2322 for(i=1; i<nCmp; i++){ 2323 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2324 ** of the index. If not, exit the loop. */ 2325 char aff; /* Comparison affinity */ 2326 char idxaff = 0; /* Indexed columns affinity */ 2327 CollSeq *pColl; /* Comparison collation sequence */ 2328 Expr *pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2329 Expr *pRhs = pTerm->pExpr->pRight; 2330 if( pRhs->flags & EP_xIsSelect ){ 2331 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2332 }else{ 2333 pRhs = pRhs->x.pList->a[i].pExpr; 2334 } 2335 2336 /* Check that the LHS of the comparison is a column reference to 2337 ** the right column of the right source table. And that the sort 2338 ** order of the index column is the same as the sort order of the 2339 ** leftmost index column. */ 2340 if( pLhs->op!=TK_COLUMN 2341 || pLhs->iTable!=iCur 2342 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2343 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2344 ){ 2345 break; 2346 } 2347 2348 testcase( pLhs->iColumn==XN_ROWID ); 2349 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2350 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2351 if( aff!=idxaff ) break; 2352 2353 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2354 if( pColl==0 ) break; 2355 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2356 } 2357 return i; 2358 } 2359 2360 /* 2361 ** Adjust the cost C by the costMult facter T. This only occurs if 2362 ** compiled with -DSQLITE_ENABLE_COSTMULT 2363 */ 2364 #ifdef SQLITE_ENABLE_COSTMULT 2365 # define ApplyCostMultiplier(C,T) C += T 2366 #else 2367 # define ApplyCostMultiplier(C,T) 2368 #endif 2369 2370 /* 2371 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2372 ** index pIndex. Try to match one more. 2373 ** 2374 ** When this function is called, pBuilder->pNew->nOut contains the 2375 ** number of rows expected to be visited by filtering using the nEq 2376 ** terms only. If it is modified, this value is restored before this 2377 ** function returns. 2378 ** 2379 ** If pProbe->tnum==0, that means pIndex is a fake index used for the 2380 ** INTEGER PRIMARY KEY. 2381 */ 2382 static int whereLoopAddBtreeIndex( 2383 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2384 struct SrcList_item *pSrc, /* FROM clause term being analyzed */ 2385 Index *pProbe, /* An index on pSrc */ 2386 LogEst nInMul /* log(Number of iterations due to IN) */ 2387 ){ 2388 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2389 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2390 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2391 WhereLoop *pNew; /* Template WhereLoop under construction */ 2392 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2393 int opMask; /* Valid operators for constraints */ 2394 WhereScan scan; /* Iterator for WHERE terms */ 2395 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2396 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2397 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2398 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2399 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2400 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2401 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2402 LogEst saved_nOut; /* Original value of pNew->nOut */ 2403 int rc = SQLITE_OK; /* Return code */ 2404 LogEst rSize; /* Number of rows in the table */ 2405 LogEst rLogSize; /* Logarithm of table size */ 2406 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2407 2408 pNew = pBuilder->pNew; 2409 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2410 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d\n", 2411 pProbe->pTable->zName,pProbe->zName, pNew->u.btree.nEq)); 2412 2413 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2414 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2415 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2416 opMask = WO_LT|WO_LE; 2417 }else{ 2418 assert( pNew->u.btree.nBtm==0 ); 2419 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2420 } 2421 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2422 2423 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2424 2425 saved_nEq = pNew->u.btree.nEq; 2426 saved_nBtm = pNew->u.btree.nBtm; 2427 saved_nTop = pNew->u.btree.nTop; 2428 saved_nSkip = pNew->nSkip; 2429 saved_nLTerm = pNew->nLTerm; 2430 saved_wsFlags = pNew->wsFlags; 2431 saved_prereq = pNew->prereq; 2432 saved_nOut = pNew->nOut; 2433 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2434 opMask, pProbe); 2435 pNew->rSetup = 0; 2436 rSize = pProbe->aiRowLogEst[0]; 2437 rLogSize = estLog(rSize); 2438 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2439 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2440 LogEst rCostIdx; 2441 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2442 int nIn = 0; 2443 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2444 int nRecValid = pBuilder->nRecValid; 2445 #endif 2446 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2447 && indexColumnNotNull(pProbe, saved_nEq) 2448 ){ 2449 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2450 } 2451 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2452 2453 /* Do not allow the upper bound of a LIKE optimization range constraint 2454 ** to mix with a lower range bound from some other source */ 2455 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2456 2457 /* Do not allow constraints from the WHERE clause to be used by the 2458 ** right table of a LEFT JOIN. Only constraints in the ON clause are 2459 ** allowed */ 2460 if( (pSrc->fg.jointype & JT_LEFT)!=0 2461 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2462 ){ 2463 continue; 2464 } 2465 2466 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2467 pBuilder->bldFlags |= SQLITE_BLDF_UNIQUE; 2468 }else{ 2469 pBuilder->bldFlags |= SQLITE_BLDF_INDEXED; 2470 } 2471 pNew->wsFlags = saved_wsFlags; 2472 pNew->u.btree.nEq = saved_nEq; 2473 pNew->u.btree.nBtm = saved_nBtm; 2474 pNew->u.btree.nTop = saved_nTop; 2475 pNew->nLTerm = saved_nLTerm; 2476 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2477 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2478 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2479 2480 assert( nInMul==0 2481 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2482 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2483 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2484 ); 2485 2486 if( eOp & WO_IN ){ 2487 Expr *pExpr = pTerm->pExpr; 2488 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 2489 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2490 int i; 2491 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2492 2493 /* The expression may actually be of the form (x, y) IN (SELECT...). 2494 ** In this case there is a separate term for each of (x) and (y). 2495 ** However, the nIn multiplier should only be applied once, not once 2496 ** for each such term. The following loop checks that pTerm is the 2497 ** first such term in use, and sets nIn back to 0 if it is not. */ 2498 for(i=0; i<pNew->nLTerm-1; i++){ 2499 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2500 } 2501 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2502 /* "x IN (value, value, ...)" */ 2503 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2504 assert( nIn>0 ); /* RHS always has 2 or more terms... The parser 2505 ** changes "x IN (?)" into "x=?". */ 2506 } 2507 if( pProbe->hasStat1 ){ 2508 LogEst M, logK, safetyMargin; 2509 /* Let: 2510 ** N = the total number of rows in the table 2511 ** K = the number of entries on the RHS of the IN operator 2512 ** M = the number of rows in the table that match terms to the 2513 ** to the left in the same index. If the IN operator is on 2514 ** the left-most index column, M==N. 2515 ** 2516 ** Given the definitions above, it is better to omit the IN operator 2517 ** from the index lookup and instead do a scan of the M elements, 2518 ** testing each scanned row against the IN operator separately, if: 2519 ** 2520 ** M*log(K) < K*log(N) 2521 ** 2522 ** Our estimates for M, K, and N might be inaccurate, so we build in 2523 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator 2524 ** with the index, as using an index has better worst-case behavior. 2525 ** If we do not have real sqlite_stat1 data, always prefer to use 2526 ** the index. 2527 */ 2528 M = pProbe->aiRowLogEst[saved_nEq]; 2529 logK = estLog(nIn); 2530 safetyMargin = 10; /* TUNING: extra weight for indexed IN */ 2531 if( M + logK + safetyMargin < nIn + rLogSize ){ 2532 WHERETRACE(0x40, 2533 ("Scan preferred over IN operator on column %d of \"%s\" (%d<%d)\n", 2534 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2535 continue; 2536 }else{ 2537 WHERETRACE(0x40, 2538 ("IN operator preferred on column %d of \"%s\" (%d>=%d)\n", 2539 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2540 } 2541 } 2542 pNew->wsFlags |= WHERE_COLUMN_IN; 2543 }else if( eOp & (WO_EQ|WO_IS) ){ 2544 int iCol = pProbe->aiColumn[saved_nEq]; 2545 pNew->wsFlags |= WHERE_COLUMN_EQ; 2546 assert( saved_nEq==pNew->u.btree.nEq ); 2547 if( iCol==XN_ROWID 2548 || (iCol>=0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2549 ){ 2550 if( iCol==XN_ROWID || pProbe->uniqNotNull 2551 || (pProbe->nKeyCol==1 && pProbe->onError && eOp==WO_EQ) 2552 ){ 2553 pNew->wsFlags |= WHERE_ONEROW; 2554 }else{ 2555 pNew->wsFlags |= WHERE_UNQ_WANTED; 2556 } 2557 } 2558 }else if( eOp & WO_ISNULL ){ 2559 pNew->wsFlags |= WHERE_COLUMN_NULL; 2560 }else if( eOp & (WO_GT|WO_GE) ){ 2561 testcase( eOp & WO_GT ); 2562 testcase( eOp & WO_GE ); 2563 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2564 pNew->u.btree.nBtm = whereRangeVectorLen( 2565 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2566 ); 2567 pBtm = pTerm; 2568 pTop = 0; 2569 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2570 /* Range contraints that come from the LIKE optimization are 2571 ** always used in pairs. */ 2572 pTop = &pTerm[1]; 2573 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2574 assert( pTop->wtFlags & TERM_LIKEOPT ); 2575 assert( pTop->eOperator==WO_LT ); 2576 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2577 pNew->aLTerm[pNew->nLTerm++] = pTop; 2578 pNew->wsFlags |= WHERE_TOP_LIMIT; 2579 pNew->u.btree.nTop = 1; 2580 } 2581 }else{ 2582 assert( eOp & (WO_LT|WO_LE) ); 2583 testcase( eOp & WO_LT ); 2584 testcase( eOp & WO_LE ); 2585 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2586 pNew->u.btree.nTop = whereRangeVectorLen( 2587 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2588 ); 2589 pTop = pTerm; 2590 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2591 pNew->aLTerm[pNew->nLTerm-2] : 0; 2592 } 2593 2594 /* At this point pNew->nOut is set to the number of rows expected to 2595 ** be visited by the index scan before considering term pTerm, or the 2596 ** values of nIn and nInMul. In other words, assuming that all 2597 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2598 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2599 assert( pNew->nOut==saved_nOut ); 2600 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2601 /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4 2602 ** data, using some other estimate. */ 2603 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2604 }else{ 2605 int nEq = ++pNew->u.btree.nEq; 2606 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2607 2608 assert( pNew->nOut==saved_nOut ); 2609 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2610 assert( (eOp & WO_IN) || nIn==0 ); 2611 testcase( eOp & WO_IN ); 2612 pNew->nOut += pTerm->truthProb; 2613 pNew->nOut -= nIn; 2614 }else{ 2615 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2616 tRowcnt nOut = 0; 2617 if( nInMul==0 2618 && pProbe->nSample 2619 && pNew->u.btree.nEq<=pProbe->nSampleCol 2620 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) 2621 && OptimizationEnabled(db, SQLITE_Stat34) 2622 ){ 2623 Expr *pExpr = pTerm->pExpr; 2624 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2625 testcase( eOp & WO_EQ ); 2626 testcase( eOp & WO_IS ); 2627 testcase( eOp & WO_ISNULL ); 2628 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2629 }else{ 2630 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2631 } 2632 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2633 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2634 if( nOut ){ 2635 pNew->nOut = sqlite3LogEst(nOut); 2636 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2637 pNew->nOut -= nIn; 2638 } 2639 } 2640 if( nOut==0 ) 2641 #endif 2642 { 2643 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2644 if( eOp & WO_ISNULL ){ 2645 /* TUNING: If there is no likelihood() value, assume that a 2646 ** "col IS NULL" expression matches twice as many rows 2647 ** as (col=?). */ 2648 pNew->nOut += 10; 2649 } 2650 } 2651 } 2652 } 2653 2654 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2655 ** it to pNew->rRun, which is currently set to the cost of the index 2656 ** seek only. Then, if this is a non-covering index, add the cost of 2657 ** visiting the rows in the main table. */ 2658 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2659 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2660 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2661 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2662 } 2663 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2664 2665 nOutUnadjusted = pNew->nOut; 2666 pNew->rRun += nInMul + nIn; 2667 pNew->nOut += nInMul + nIn; 2668 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 2669 rc = whereLoopInsert(pBuilder, pNew); 2670 2671 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2672 pNew->nOut = saved_nOut; 2673 }else{ 2674 pNew->nOut = nOutUnadjusted; 2675 } 2676 2677 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 2678 && pNew->u.btree.nEq<pProbe->nColumn 2679 ){ 2680 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 2681 } 2682 pNew->nOut = saved_nOut; 2683 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 2684 pBuilder->nRecValid = nRecValid; 2685 #endif 2686 } 2687 pNew->prereq = saved_prereq; 2688 pNew->u.btree.nEq = saved_nEq; 2689 pNew->u.btree.nBtm = saved_nBtm; 2690 pNew->u.btree.nTop = saved_nTop; 2691 pNew->nSkip = saved_nSkip; 2692 pNew->wsFlags = saved_wsFlags; 2693 pNew->nOut = saved_nOut; 2694 pNew->nLTerm = saved_nLTerm; 2695 2696 /* Consider using a skip-scan if there are no WHERE clause constraints 2697 ** available for the left-most terms of the index, and if the average 2698 ** number of repeats in the left-most terms is at least 18. 2699 ** 2700 ** The magic number 18 is selected on the basis that scanning 17 rows 2701 ** is almost always quicker than an index seek (even though if the index 2702 ** contains fewer than 2^17 rows we assume otherwise in other parts of 2703 ** the code). And, even if it is not, it should not be too much slower. 2704 ** On the other hand, the extra seeks could end up being significantly 2705 ** more expensive. */ 2706 assert( 42==sqlite3LogEst(18) ); 2707 if( saved_nEq==saved_nSkip 2708 && saved_nEq+1<pProbe->nKeyCol 2709 && pProbe->noSkipScan==0 2710 && OptimizationEnabled(db, SQLITE_SkipScan) 2711 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 2712 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 2713 ){ 2714 LogEst nIter; 2715 pNew->u.btree.nEq++; 2716 pNew->nSkip++; 2717 pNew->aLTerm[pNew->nLTerm++] = 0; 2718 pNew->wsFlags |= WHERE_SKIPSCAN; 2719 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 2720 pNew->nOut -= nIter; 2721 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 2722 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 2723 nIter += 5; 2724 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 2725 pNew->nOut = saved_nOut; 2726 pNew->u.btree.nEq = saved_nEq; 2727 pNew->nSkip = saved_nSkip; 2728 pNew->wsFlags = saved_wsFlags; 2729 } 2730 2731 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n", 2732 pProbe->pTable->zName, pProbe->zName, saved_nEq, rc)); 2733 return rc; 2734 } 2735 2736 /* 2737 ** Return True if it is possible that pIndex might be useful in 2738 ** implementing the ORDER BY clause in pBuilder. 2739 ** 2740 ** Return False if pBuilder does not contain an ORDER BY clause or 2741 ** if there is no way for pIndex to be useful in implementing that 2742 ** ORDER BY clause. 2743 */ 2744 static int indexMightHelpWithOrderBy( 2745 WhereLoopBuilder *pBuilder, 2746 Index *pIndex, 2747 int iCursor 2748 ){ 2749 ExprList *pOB; 2750 ExprList *aColExpr; 2751 int ii, jj; 2752 2753 if( pIndex->bUnordered ) return 0; 2754 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 2755 for(ii=0; ii<pOB->nExpr; ii++){ 2756 Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr); 2757 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 2758 if( pExpr->iColumn<0 ) return 1; 2759 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2760 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 2761 } 2762 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 2763 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2764 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 2765 if( sqlite3ExprCompareSkip(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 2766 return 1; 2767 } 2768 } 2769 } 2770 } 2771 return 0; 2772 } 2773 2774 /* Check to see if a partial index with pPartIndexWhere can be used 2775 ** in the current query. Return true if it can be and false if not. 2776 */ 2777 static int whereUsablePartialIndex(int iTab, WhereClause *pWC, Expr *pWhere){ 2778 int i; 2779 WhereTerm *pTerm; 2780 Parse *pParse = pWC->pWInfo->pParse; 2781 while( pWhere->op==TK_AND ){ 2782 if( !whereUsablePartialIndex(iTab,pWC,pWhere->pLeft) ) return 0; 2783 pWhere = pWhere->pRight; 2784 } 2785 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 2786 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2787 Expr *pExpr = pTerm->pExpr; 2788 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 2789 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 2790 ){ 2791 return 1; 2792 } 2793 } 2794 return 0; 2795 } 2796 2797 /* 2798 ** Add all WhereLoop objects for a single table of the join where the table 2799 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 2800 ** a b-tree table, not a virtual table. 2801 ** 2802 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 2803 ** are calculated as follows: 2804 ** 2805 ** For a full scan, assuming the table (or index) contains nRow rows: 2806 ** 2807 ** cost = nRow * 3.0 // full-table scan 2808 ** cost = nRow * K // scan of covering index 2809 ** cost = nRow * (K+3.0) // scan of non-covering index 2810 ** 2811 ** where K is a value between 1.1 and 3.0 set based on the relative 2812 ** estimated average size of the index and table records. 2813 ** 2814 ** For an index scan, where nVisit is the number of index rows visited 2815 ** by the scan, and nSeek is the number of seek operations required on 2816 ** the index b-tree: 2817 ** 2818 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 2819 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 2820 ** 2821 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 2822 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 2823 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 2824 ** 2825 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 2826 ** of uncertainty. For this reason, scoring is designed to pick plans that 2827 ** "do the least harm" if the estimates are inaccurate. For example, a 2828 ** log(nRow) factor is omitted from a non-covering index scan in order to 2829 ** bias the scoring in favor of using an index, since the worst-case 2830 ** performance of using an index is far better than the worst-case performance 2831 ** of a full table scan. 2832 */ 2833 static int whereLoopAddBtree( 2834 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2835 Bitmask mPrereq /* Extra prerequesites for using this table */ 2836 ){ 2837 WhereInfo *pWInfo; /* WHERE analysis context */ 2838 Index *pProbe; /* An index we are evaluating */ 2839 Index sPk; /* A fake index object for the primary key */ 2840 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 2841 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 2842 SrcList *pTabList; /* The FROM clause */ 2843 struct SrcList_item *pSrc; /* The FROM clause btree term to add */ 2844 WhereLoop *pNew; /* Template WhereLoop object */ 2845 int rc = SQLITE_OK; /* Return code */ 2846 int iSortIdx = 1; /* Index number */ 2847 int b; /* A boolean value */ 2848 LogEst rSize; /* number of rows in the table */ 2849 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 2850 WhereClause *pWC; /* The parsed WHERE clause */ 2851 Table *pTab; /* Table being queried */ 2852 2853 pNew = pBuilder->pNew; 2854 pWInfo = pBuilder->pWInfo; 2855 pTabList = pWInfo->pTabList; 2856 pSrc = pTabList->a + pNew->iTab; 2857 pTab = pSrc->pTab; 2858 pWC = pBuilder->pWC; 2859 assert( !IsVirtual(pSrc->pTab) ); 2860 2861 if( pSrc->pIBIndex ){ 2862 /* An INDEXED BY clause specifies a particular index to use */ 2863 pProbe = pSrc->pIBIndex; 2864 }else if( !HasRowid(pTab) ){ 2865 pProbe = pTab->pIndex; 2866 }else{ 2867 /* There is no INDEXED BY clause. Create a fake Index object in local 2868 ** variable sPk to represent the rowid primary key index. Make this 2869 ** fake index the first in a chain of Index objects with all of the real 2870 ** indices to follow */ 2871 Index *pFirst; /* First of real indices on the table */ 2872 memset(&sPk, 0, sizeof(Index)); 2873 sPk.nKeyCol = 1; 2874 sPk.nColumn = 1; 2875 sPk.aiColumn = &aiColumnPk; 2876 sPk.aiRowLogEst = aiRowEstPk; 2877 sPk.onError = OE_Replace; 2878 sPk.pTable = pTab; 2879 sPk.szIdxRow = pTab->szTabRow; 2880 aiRowEstPk[0] = pTab->nRowLogEst; 2881 aiRowEstPk[1] = 0; 2882 pFirst = pSrc->pTab->pIndex; 2883 if( pSrc->fg.notIndexed==0 ){ 2884 /* The real indices of the table are only considered if the 2885 ** NOT INDEXED qualifier is omitted from the FROM clause */ 2886 sPk.pNext = pFirst; 2887 } 2888 pProbe = &sPk; 2889 } 2890 rSize = pTab->nRowLogEst; 2891 rLogSize = estLog(rSize); 2892 2893 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 2894 /* Automatic indexes */ 2895 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 2896 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 2897 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 2898 && pSrc->pIBIndex==0 /* Has no INDEXED BY clause */ 2899 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 2900 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 2901 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 2902 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 2903 ){ 2904 /* Generate auto-index WhereLoops */ 2905 WhereTerm *pTerm; 2906 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 2907 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 2908 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2909 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 2910 pNew->u.btree.nEq = 1; 2911 pNew->nSkip = 0; 2912 pNew->u.btree.pIndex = 0; 2913 pNew->nLTerm = 1; 2914 pNew->aLTerm[0] = pTerm; 2915 /* TUNING: One-time cost for computing the automatic index is 2916 ** estimated to be X*N*log2(N) where N is the number of rows in 2917 ** the table being indexed and where X is 7 (LogEst=28) for normal 2918 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value 2919 ** of X is smaller for views and subqueries so that the query planner 2920 ** will be more aggressive about generating automatic indexes for 2921 ** those objects, since there is no opportunity to add schema 2922 ** indexes on subqueries and views. */ 2923 pNew->rSetup = rLogSize + rSize; 2924 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ 2925 pNew->rSetup += 28; 2926 }else{ 2927 pNew->rSetup -= 10; 2928 } 2929 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 2930 if( pNew->rSetup<0 ) pNew->rSetup = 0; 2931 /* TUNING: Each index lookup yields 20 rows in the table. This 2932 ** is more than the usual guess of 10 rows, since we have no way 2933 ** of knowing how selective the index will ultimately be. It would 2934 ** not be unreasonable to make this value much larger. */ 2935 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 2936 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 2937 pNew->wsFlags = WHERE_AUTO_INDEX; 2938 pNew->prereq = mPrereq | pTerm->prereqRight; 2939 rc = whereLoopInsert(pBuilder, pNew); 2940 } 2941 } 2942 } 2943 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 2944 2945 /* Loop over all indices. If there was an INDEXED BY clause, then only 2946 ** consider index pProbe. */ 2947 for(; rc==SQLITE_OK && pProbe; 2948 pProbe=(pSrc->pIBIndex ? 0 : pProbe->pNext), iSortIdx++ 2949 ){ 2950 if( pProbe->pPartIdxWhere!=0 2951 && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){ 2952 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 2953 continue; /* Partial index inappropriate for this query */ 2954 } 2955 if( pProbe->bNoQuery ) continue; 2956 rSize = pProbe->aiRowLogEst[0]; 2957 pNew->u.btree.nEq = 0; 2958 pNew->u.btree.nBtm = 0; 2959 pNew->u.btree.nTop = 0; 2960 pNew->nSkip = 0; 2961 pNew->nLTerm = 0; 2962 pNew->iSortIdx = 0; 2963 pNew->rSetup = 0; 2964 pNew->prereq = mPrereq; 2965 pNew->nOut = rSize; 2966 pNew->u.btree.pIndex = pProbe; 2967 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 2968 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 2969 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 2970 if( pProbe->tnum<=0 ){ 2971 /* Integer primary key index */ 2972 pNew->wsFlags = WHERE_IPK; 2973 2974 /* Full table scan */ 2975 pNew->iSortIdx = b ? iSortIdx : 0; 2976 /* TUNING: Cost of full table scan is (N*3.0). */ 2977 pNew->rRun = rSize + 16; 2978 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 2979 whereLoopOutputAdjust(pWC, pNew, rSize); 2980 rc = whereLoopInsert(pBuilder, pNew); 2981 pNew->nOut = rSize; 2982 if( rc ) break; 2983 }else{ 2984 Bitmask m; 2985 if( pProbe->isCovering ){ 2986 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 2987 m = 0; 2988 }else{ 2989 m = pSrc->colUsed & pProbe->colNotIdxed; 2990 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 2991 } 2992 2993 /* Full scan via index */ 2994 if( b 2995 || !HasRowid(pTab) 2996 || pProbe->pPartIdxWhere!=0 2997 || ( m==0 2998 && pProbe->bUnordered==0 2999 && (pProbe->szIdxRow<pTab->szTabRow) 3000 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 3001 && sqlite3GlobalConfig.bUseCis 3002 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 3003 ) 3004 ){ 3005 pNew->iSortIdx = b ? iSortIdx : 0; 3006 3007 /* The cost of visiting the index rows is N*K, where K is 3008 ** between 1.1 and 3.0, depending on the relative sizes of the 3009 ** index and table rows. */ 3010 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 3011 if( m!=0 ){ 3012 /* If this is a non-covering index scan, add in the cost of 3013 ** doing table lookups. The cost will be 3x the number of 3014 ** lookups. Take into account WHERE clause terms that can be 3015 ** satisfied using just the index, and that do not require a 3016 ** table lookup. */ 3017 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 3018 int ii; 3019 int iCur = pSrc->iCursor; 3020 WhereClause *pWC2 = &pWInfo->sWC; 3021 for(ii=0; ii<pWC2->nTerm; ii++){ 3022 WhereTerm *pTerm = &pWC2->a[ii]; 3023 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 3024 break; 3025 } 3026 /* pTerm can be evaluated using just the index. So reduce 3027 ** the expected number of table lookups accordingly */ 3028 if( pTerm->truthProb<=0 ){ 3029 nLookup += pTerm->truthProb; 3030 }else{ 3031 nLookup--; 3032 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 3033 } 3034 } 3035 3036 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 3037 } 3038 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3039 whereLoopOutputAdjust(pWC, pNew, rSize); 3040 rc = whereLoopInsert(pBuilder, pNew); 3041 pNew->nOut = rSize; 3042 if( rc ) break; 3043 } 3044 } 3045 3046 pBuilder->bldFlags = 0; 3047 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 3048 if( pBuilder->bldFlags==SQLITE_BLDF_INDEXED ){ 3049 /* If a non-unique index is used, or if a prefix of the key for 3050 ** unique index is used (making the index functionally non-unique) 3051 ** then the sqlite_stat1 data becomes important for scoring the 3052 ** plan */ 3053 pTab->tabFlags |= TF_StatsUsed; 3054 } 3055 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 3056 sqlite3Stat4ProbeFree(pBuilder->pRec); 3057 pBuilder->nRecValid = 0; 3058 pBuilder->pRec = 0; 3059 #endif 3060 } 3061 return rc; 3062 } 3063 3064 #ifndef SQLITE_OMIT_VIRTUALTABLE 3065 3066 /* 3067 ** Argument pIdxInfo is already populated with all constraints that may 3068 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 3069 ** function marks a subset of those constraints usable, invokes the 3070 ** xBestIndex method and adds the returned plan to pBuilder. 3071 ** 3072 ** A constraint is marked usable if: 3073 ** 3074 ** * Argument mUsable indicates that its prerequisites are available, and 3075 ** 3076 ** * It is not one of the operators specified in the mExclude mask passed 3077 ** as the fourth argument (which in practice is either WO_IN or 0). 3078 ** 3079 ** Argument mPrereq is a mask of tables that must be scanned before the 3080 ** virtual table in question. These are added to the plans prerequisites 3081 ** before it is added to pBuilder. 3082 ** 3083 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3084 ** uses one or more WO_IN terms, or false otherwise. 3085 */ 3086 static int whereLoopAddVirtualOne( 3087 WhereLoopBuilder *pBuilder, 3088 Bitmask mPrereq, /* Mask of tables that must be used. */ 3089 Bitmask mUsable, /* Mask of usable tables */ 3090 u16 mExclude, /* Exclude terms using these operators */ 3091 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3092 u16 mNoOmit, /* Do not omit these constraints */ 3093 int *pbIn /* OUT: True if plan uses an IN(...) op */ 3094 ){ 3095 WhereClause *pWC = pBuilder->pWC; 3096 struct sqlite3_index_constraint *pIdxCons; 3097 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3098 int i; 3099 int mxTerm; 3100 int rc = SQLITE_OK; 3101 WhereLoop *pNew = pBuilder->pNew; 3102 Parse *pParse = pBuilder->pWInfo->pParse; 3103 struct SrcList_item *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3104 int nConstraint = pIdxInfo->nConstraint; 3105 3106 assert( (mUsable & mPrereq)==mPrereq ); 3107 *pbIn = 0; 3108 pNew->prereq = mPrereq; 3109 3110 /* Set the usable flag on the subset of constraints identified by 3111 ** arguments mUsable and mExclude. */ 3112 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3113 for(i=0; i<nConstraint; i++, pIdxCons++){ 3114 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3115 pIdxCons->usable = 0; 3116 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3117 && (pTerm->eOperator & mExclude)==0 3118 ){ 3119 pIdxCons->usable = 1; 3120 } 3121 } 3122 3123 /* Initialize the output fields of the sqlite3_index_info structure */ 3124 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3125 assert( pIdxInfo->needToFreeIdxStr==0 ); 3126 pIdxInfo->idxStr = 0; 3127 pIdxInfo->idxNum = 0; 3128 pIdxInfo->orderByConsumed = 0; 3129 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3130 pIdxInfo->estimatedRows = 25; 3131 pIdxInfo->idxFlags = 0; 3132 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3133 3134 /* Invoke the virtual table xBestIndex() method */ 3135 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3136 if( rc ){ 3137 if( rc==SQLITE_CONSTRAINT ){ 3138 /* If the xBestIndex method returns SQLITE_CONSTRAINT, that means 3139 ** that the particular combination of parameters provided is unusable. 3140 ** Make no entries in the loop table. 3141 */ 3142 WHERETRACE(0xffff, (" ^^^^--- non-viable plan rejected!\n")); 3143 return SQLITE_OK; 3144 } 3145 return rc; 3146 } 3147 3148 mxTerm = -1; 3149 assert( pNew->nLSlot>=nConstraint ); 3150 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 3151 pNew->u.vtab.omitMask = 0; 3152 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3153 for(i=0; i<nConstraint; i++, pIdxCons++){ 3154 int iTerm; 3155 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3156 WhereTerm *pTerm; 3157 int j = pIdxCons->iTermOffset; 3158 if( iTerm>=nConstraint 3159 || j<0 3160 || j>=pWC->nTerm 3161 || pNew->aLTerm[iTerm]!=0 3162 || pIdxCons->usable==0 3163 ){ 3164 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3165 testcase( pIdxInfo->needToFreeIdxStr ); 3166 return SQLITE_ERROR; 3167 } 3168 testcase( iTerm==nConstraint-1 ); 3169 testcase( j==0 ); 3170 testcase( j==pWC->nTerm-1 ); 3171 pTerm = &pWC->a[j]; 3172 pNew->prereq |= pTerm->prereqRight; 3173 assert( iTerm<pNew->nLSlot ); 3174 pNew->aLTerm[iTerm] = pTerm; 3175 if( iTerm>mxTerm ) mxTerm = iTerm; 3176 testcase( iTerm==15 ); 3177 testcase( iTerm==16 ); 3178 if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<<iTerm; 3179 if( (pTerm->eOperator & WO_IN)!=0 ){ 3180 /* A virtual table that is constrained by an IN clause may not 3181 ** consume the ORDER BY clause because (1) the order of IN terms 3182 ** is not necessarily related to the order of output terms and 3183 ** (2) Multiple outputs from a single IN value will not merge 3184 ** together. */ 3185 pIdxInfo->orderByConsumed = 0; 3186 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3187 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3188 } 3189 } 3190 } 3191 pNew->u.vtab.omitMask &= ~mNoOmit; 3192 3193 pNew->nLTerm = mxTerm+1; 3194 for(i=0; i<=mxTerm; i++){ 3195 if( pNew->aLTerm[i]==0 ){ 3196 /* The non-zero argvIdx values must be contiguous. Raise an 3197 ** error if they are not */ 3198 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3199 testcase( pIdxInfo->needToFreeIdxStr ); 3200 return SQLITE_ERROR; 3201 } 3202 } 3203 assert( pNew->nLTerm<=pNew->nLSlot ); 3204 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3205 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3206 pIdxInfo->needToFreeIdxStr = 0; 3207 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3208 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3209 pIdxInfo->nOrderBy : 0); 3210 pNew->rSetup = 0; 3211 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3212 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3213 3214 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3215 ** that the scan will visit at most one row. Clear it otherwise. */ 3216 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3217 pNew->wsFlags |= WHERE_ONEROW; 3218 }else{ 3219 pNew->wsFlags &= ~WHERE_ONEROW; 3220 } 3221 rc = whereLoopInsert(pBuilder, pNew); 3222 if( pNew->u.vtab.needFree ){ 3223 sqlite3_free(pNew->u.vtab.idxStr); 3224 pNew->u.vtab.needFree = 0; 3225 } 3226 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3227 *pbIn, (sqlite3_uint64)mPrereq, 3228 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3229 3230 return rc; 3231 } 3232 3233 /* 3234 ** If this function is invoked from within an xBestIndex() callback, it 3235 ** returns a pointer to a buffer containing the name of the collation 3236 ** sequence associated with element iCons of the sqlite3_index_info.aConstraint 3237 ** array. Or, if iCons is out of range or there is no active xBestIndex 3238 ** call, return NULL. 3239 */ 3240 const char *sqlite3_vtab_collation(sqlite3_index_info *pIdxInfo, int iCons){ 3241 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3242 const char *zRet = 0; 3243 if( iCons>=0 && iCons<pIdxInfo->nConstraint ){ 3244 CollSeq *pC = 0; 3245 int iTerm = pIdxInfo->aConstraint[iCons].iTermOffset; 3246 Expr *pX = pHidden->pWC->a[iTerm].pExpr; 3247 if( pX->pLeft ){ 3248 pC = sqlite3BinaryCompareCollSeq(pHidden->pParse, pX->pLeft, pX->pRight); 3249 } 3250 zRet = (pC ? pC->zName : sqlite3StrBINARY); 3251 } 3252 return zRet; 3253 } 3254 3255 /* 3256 ** Add all WhereLoop objects for a table of the join identified by 3257 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3258 ** 3259 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3260 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3261 ** entries that occur before the virtual table in the FROM clause and are 3262 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3263 ** mUnusable mask contains all FROM clause entries that occur after the 3264 ** virtual table and are separated from it by at least one LEFT or 3265 ** CROSS JOIN. 3266 ** 3267 ** For example, if the query were: 3268 ** 3269 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3270 ** 3271 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3272 ** 3273 ** All the tables in mPrereq must be scanned before the current virtual 3274 ** table. So any terms for which all prerequisites are satisfied by 3275 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3276 ** Conversely, all tables in mUnusable must be scanned after the current 3277 ** virtual table, so any terms for which the prerequisites overlap with 3278 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3279 */ 3280 static int whereLoopAddVirtual( 3281 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3282 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3283 Bitmask mUnusable /* Tables that must be scanned after this one */ 3284 ){ 3285 int rc = SQLITE_OK; /* Return code */ 3286 WhereInfo *pWInfo; /* WHERE analysis context */ 3287 Parse *pParse; /* The parsing context */ 3288 WhereClause *pWC; /* The WHERE clause */ 3289 struct SrcList_item *pSrc; /* The FROM clause term to search */ 3290 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3291 int nConstraint; /* Number of constraints in p */ 3292 int bIn; /* True if plan uses IN(...) operator */ 3293 WhereLoop *pNew; 3294 Bitmask mBest; /* Tables used by best possible plan */ 3295 u16 mNoOmit; 3296 3297 assert( (mPrereq & mUnusable)==0 ); 3298 pWInfo = pBuilder->pWInfo; 3299 pParse = pWInfo->pParse; 3300 pWC = pBuilder->pWC; 3301 pNew = pBuilder->pNew; 3302 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3303 assert( IsVirtual(pSrc->pTab) ); 3304 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy, 3305 &mNoOmit); 3306 if( p==0 ) return SQLITE_NOMEM_BKPT; 3307 pNew->rSetup = 0; 3308 pNew->wsFlags = WHERE_VIRTUALTABLE; 3309 pNew->nLTerm = 0; 3310 pNew->u.vtab.needFree = 0; 3311 nConstraint = p->nConstraint; 3312 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3313 sqlite3DbFree(pParse->db, p); 3314 return SQLITE_NOMEM_BKPT; 3315 } 3316 3317 /* First call xBestIndex() with all constraints usable. */ 3318 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc->pTab->zName)); 3319 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3320 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn); 3321 3322 /* If the call to xBestIndex() with all terms enabled produced a plan 3323 ** that does not require any source tables (IOW: a plan with mBest==0), 3324 ** then there is no point in making any further calls to xBestIndex() 3325 ** since they will all return the same result (if the xBestIndex() 3326 ** implementation is sane). */ 3327 if( rc==SQLITE_OK && (mBest = (pNew->prereq & ~mPrereq))!=0 ){ 3328 int seenZero = 0; /* True if a plan with no prereqs seen */ 3329 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3330 Bitmask mPrev = 0; 3331 Bitmask mBestNoIn = 0; 3332 3333 /* If the plan produced by the earlier call uses an IN(...) term, call 3334 ** xBestIndex again, this time with IN(...) terms disabled. */ 3335 if( bIn ){ 3336 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3337 rc = whereLoopAddVirtualOne( 3338 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn); 3339 assert( bIn==0 ); 3340 mBestNoIn = pNew->prereq & ~mPrereq; 3341 if( mBestNoIn==0 ){ 3342 seenZero = 1; 3343 seenZeroNoIN = 1; 3344 } 3345 } 3346 3347 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3348 ** in the set of terms that apply to the current virtual table. */ 3349 while( rc==SQLITE_OK ){ 3350 int i; 3351 Bitmask mNext = ALLBITS; 3352 assert( mNext>0 ); 3353 for(i=0; i<nConstraint; i++){ 3354 Bitmask mThis = ( 3355 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3356 ); 3357 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3358 } 3359 mPrev = mNext; 3360 if( mNext==ALLBITS ) break; 3361 if( mNext==mBest || mNext==mBestNoIn ) continue; 3362 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3363 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3364 rc = whereLoopAddVirtualOne( 3365 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn); 3366 if( pNew->prereq==mPrereq ){ 3367 seenZero = 1; 3368 if( bIn==0 ) seenZeroNoIN = 1; 3369 } 3370 } 3371 3372 /* If the calls to xBestIndex() in the above loop did not find a plan 3373 ** that requires no source tables at all (i.e. one guaranteed to be 3374 ** usable), make a call here with all source tables disabled */ 3375 if( rc==SQLITE_OK && seenZero==0 ){ 3376 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3377 rc = whereLoopAddVirtualOne( 3378 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn); 3379 if( bIn==0 ) seenZeroNoIN = 1; 3380 } 3381 3382 /* If the calls to xBestIndex() have so far failed to find a plan 3383 ** that requires no source tables at all and does not use an IN(...) 3384 ** operator, make a final call to obtain one here. */ 3385 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3386 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3387 rc = whereLoopAddVirtualOne( 3388 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn); 3389 } 3390 } 3391 3392 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3393 sqlite3DbFreeNN(pParse->db, p); 3394 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc->pTab->zName, rc)); 3395 return rc; 3396 } 3397 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3398 3399 /* 3400 ** Add WhereLoop entries to handle OR terms. This works for either 3401 ** btrees or virtual tables. 3402 */ 3403 static int whereLoopAddOr( 3404 WhereLoopBuilder *pBuilder, 3405 Bitmask mPrereq, 3406 Bitmask mUnusable 3407 ){ 3408 WhereInfo *pWInfo = pBuilder->pWInfo; 3409 WhereClause *pWC; 3410 WhereLoop *pNew; 3411 WhereTerm *pTerm, *pWCEnd; 3412 int rc = SQLITE_OK; 3413 int iCur; 3414 WhereClause tempWC; 3415 WhereLoopBuilder sSubBuild; 3416 WhereOrSet sSum, sCur; 3417 struct SrcList_item *pItem; 3418 3419 pWC = pBuilder->pWC; 3420 pWCEnd = pWC->a + pWC->nTerm; 3421 pNew = pBuilder->pNew; 3422 memset(&sSum, 0, sizeof(sSum)); 3423 pItem = pWInfo->pTabList->a + pNew->iTab; 3424 iCur = pItem->iCursor; 3425 3426 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3427 if( (pTerm->eOperator & WO_OR)!=0 3428 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3429 ){ 3430 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3431 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3432 WhereTerm *pOrTerm; 3433 int once = 1; 3434 int i, j; 3435 3436 sSubBuild = *pBuilder; 3437 sSubBuild.pOrderBy = 0; 3438 sSubBuild.pOrSet = &sCur; 3439 3440 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3441 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3442 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3443 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3444 }else if( pOrTerm->leftCursor==iCur ){ 3445 tempWC.pWInfo = pWC->pWInfo; 3446 tempWC.pOuter = pWC; 3447 tempWC.op = TK_AND; 3448 tempWC.nTerm = 1; 3449 tempWC.a = pOrTerm; 3450 sSubBuild.pWC = &tempWC; 3451 }else{ 3452 continue; 3453 } 3454 sCur.n = 0; 3455 #ifdef WHERETRACE_ENABLED 3456 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3457 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3458 if( sqlite3WhereTrace & 0x400 ){ 3459 sqlite3WhereClausePrint(sSubBuild.pWC); 3460 } 3461 #endif 3462 #ifndef SQLITE_OMIT_VIRTUALTABLE 3463 if( IsVirtual(pItem->pTab) ){ 3464 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 3465 }else 3466 #endif 3467 { 3468 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 3469 } 3470 if( rc==SQLITE_OK ){ 3471 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 3472 } 3473 assert( rc==SQLITE_OK || sCur.n==0 ); 3474 if( sCur.n==0 ){ 3475 sSum.n = 0; 3476 break; 3477 }else if( once ){ 3478 whereOrMove(&sSum, &sCur); 3479 once = 0; 3480 }else{ 3481 WhereOrSet sPrev; 3482 whereOrMove(&sPrev, &sSum); 3483 sSum.n = 0; 3484 for(i=0; i<sPrev.n; i++){ 3485 for(j=0; j<sCur.n; j++){ 3486 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3487 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3488 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3489 } 3490 } 3491 } 3492 } 3493 pNew->nLTerm = 1; 3494 pNew->aLTerm[0] = pTerm; 3495 pNew->wsFlags = WHERE_MULTI_OR; 3496 pNew->rSetup = 0; 3497 pNew->iSortIdx = 0; 3498 memset(&pNew->u, 0, sizeof(pNew->u)); 3499 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3500 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3501 ** of all sub-scans required by the OR-scan. However, due to rounding 3502 ** errors, it may be that the cost of the OR-scan is equal to its 3503 ** most expensive sub-scan. Add the smallest possible penalty 3504 ** (equivalent to multiplying the cost by 1.07) to ensure that 3505 ** this does not happen. Otherwise, for WHERE clauses such as the 3506 ** following where there is an index on "y": 3507 ** 3508 ** WHERE likelihood(x=?, 0.99) OR y=? 3509 ** 3510 ** the planner may elect to "OR" together a full-table scan and an 3511 ** index lookup. And other similarly odd results. */ 3512 pNew->rRun = sSum.a[i].rRun + 1; 3513 pNew->nOut = sSum.a[i].nOut; 3514 pNew->prereq = sSum.a[i].prereq; 3515 rc = whereLoopInsert(pBuilder, pNew); 3516 } 3517 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3518 } 3519 } 3520 return rc; 3521 } 3522 3523 /* 3524 ** Add all WhereLoop objects for all tables 3525 */ 3526 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3527 WhereInfo *pWInfo = pBuilder->pWInfo; 3528 Bitmask mPrereq = 0; 3529 Bitmask mPrior = 0; 3530 int iTab; 3531 SrcList *pTabList = pWInfo->pTabList; 3532 struct SrcList_item *pItem; 3533 struct SrcList_item *pEnd = &pTabList->a[pWInfo->nLevel]; 3534 sqlite3 *db = pWInfo->pParse->db; 3535 int rc = SQLITE_OK; 3536 WhereLoop *pNew; 3537 u8 priorJointype = 0; 3538 3539 /* Loop over the tables in the join, from left to right */ 3540 pNew = pBuilder->pNew; 3541 whereLoopInit(pNew); 3542 pBuilder->iPlanLimit = SQLITE_QUERY_PLANNER_LIMIT; 3543 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3544 Bitmask mUnusable = 0; 3545 pNew->iTab = iTab; 3546 pBuilder->iPlanLimit += SQLITE_QUERY_PLANNER_LIMIT_INCR; 3547 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3548 if( ((pItem->fg.jointype|priorJointype) & (JT_LEFT|JT_CROSS))!=0 ){ 3549 /* This condition is true when pItem is the FROM clause term on the 3550 ** right-hand-side of a LEFT or CROSS JOIN. */ 3551 mPrereq = mPrior; 3552 } 3553 priorJointype = pItem->fg.jointype; 3554 #ifndef SQLITE_OMIT_VIRTUALTABLE 3555 if( IsVirtual(pItem->pTab) ){ 3556 struct SrcList_item *p; 3557 for(p=&pItem[1]; p<pEnd; p++){ 3558 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3559 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3560 } 3561 } 3562 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 3563 }else 3564 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3565 { 3566 rc = whereLoopAddBtree(pBuilder, mPrereq); 3567 } 3568 if( rc==SQLITE_OK && pBuilder->pWC->hasOr ){ 3569 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 3570 } 3571 mPrior |= pNew->maskSelf; 3572 if( rc || db->mallocFailed ){ 3573 if( rc==SQLITE_DONE ){ 3574 /* We hit the query planner search limit set by iPlanLimit */ 3575 sqlite3_log(SQLITE_WARNING, "abbreviated query algorithm search"); 3576 rc = SQLITE_OK; 3577 }else{ 3578 break; 3579 } 3580 } 3581 } 3582 3583 whereLoopClear(db, pNew); 3584 return rc; 3585 } 3586 3587 /* 3588 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 3589 ** parameters) to see if it outputs rows in the requested ORDER BY 3590 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3591 ** 3592 ** N>0: N terms of the ORDER BY clause are satisfied 3593 ** N==0: No terms of the ORDER BY clause are satisfied 3594 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3595 ** 3596 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3597 ** strict. With GROUP BY and DISTINCT the only requirement is that 3598 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3599 ** and DISTINCT do not require rows to appear in any particular order as long 3600 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3601 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3602 ** pOrderBy terms must be matched in strict left-to-right order. 3603 */ 3604 static i8 wherePathSatisfiesOrderBy( 3605 WhereInfo *pWInfo, /* The WHERE clause */ 3606 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3607 WherePath *pPath, /* The WherePath to check */ 3608 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 3609 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3610 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3611 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3612 ){ 3613 u8 revSet; /* True if rev is known */ 3614 u8 rev; /* Composite sort order */ 3615 u8 revIdx; /* Index sort order */ 3616 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 3617 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 3618 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 3619 u16 eqOpMask; /* Allowed equality operators */ 3620 u16 nKeyCol; /* Number of key columns in pIndex */ 3621 u16 nColumn; /* Total number of ordered columns in the index */ 3622 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 3623 int iLoop; /* Index of WhereLoop in pPath being processed */ 3624 int i, j; /* Loop counters */ 3625 int iCur; /* Cursor number for current WhereLoop */ 3626 int iColumn; /* A column number within table iCur */ 3627 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 3628 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3629 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 3630 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 3631 Index *pIndex; /* The index associated with pLoop */ 3632 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 3633 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 3634 Bitmask obDone; /* Mask of all ORDER BY terms */ 3635 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 3636 Bitmask ready; /* Mask of inner loops */ 3637 3638 /* 3639 ** We say the WhereLoop is "one-row" if it generates no more than one 3640 ** row of output. A WhereLoop is one-row if all of the following are true: 3641 ** (a) All index columns match with WHERE_COLUMN_EQ. 3642 ** (b) The index is unique 3643 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 3644 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 3645 ** 3646 ** We say the WhereLoop is "order-distinct" if the set of columns from 3647 ** that WhereLoop that are in the ORDER BY clause are different for every 3648 ** row of the WhereLoop. Every one-row WhereLoop is automatically 3649 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 3650 ** is not order-distinct. To be order-distinct is not quite the same as being 3651 ** UNIQUE since a UNIQUE column or index can have multiple rows that 3652 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 3653 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 3654 ** 3655 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 3656 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 3657 ** automatically order-distinct. 3658 */ 3659 3660 assert( pOrderBy!=0 ); 3661 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 3662 3663 nOrderBy = pOrderBy->nExpr; 3664 testcase( nOrderBy==BMS-1 ); 3665 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 3666 isOrderDistinct = 1; 3667 obDone = MASKBIT(nOrderBy)-1; 3668 orderDistinctMask = 0; 3669 ready = 0; 3670 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 3671 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) eqOpMask |= WO_IN; 3672 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 3673 if( iLoop>0 ) ready |= pLoop->maskSelf; 3674 if( iLoop<nLoop ){ 3675 pLoop = pPath->aLoop[iLoop]; 3676 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 3677 }else{ 3678 pLoop = pLast; 3679 } 3680 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 3681 if( pLoop->u.vtab.isOrdered ) obSat = obDone; 3682 break; 3683 }else{ 3684 pLoop->u.btree.nIdxCol = 0; 3685 } 3686 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 3687 3688 /* Mark off any ORDER BY term X that is a column in the table of 3689 ** the current loop for which there is term in the WHERE 3690 ** clause of the form X IS NULL or X=? that reference only outer 3691 ** loops. 3692 */ 3693 for(i=0; i<nOrderBy; i++){ 3694 if( MASKBIT(i) & obSat ) continue; 3695 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3696 if( pOBExpr->op!=TK_COLUMN ) continue; 3697 if( pOBExpr->iTable!=iCur ) continue; 3698 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 3699 ~ready, eqOpMask, 0); 3700 if( pTerm==0 ) continue; 3701 if( pTerm->eOperator==WO_IN ){ 3702 /* IN terms are only valid for sorting in the ORDER BY LIMIT 3703 ** optimization, and then only if they are actually used 3704 ** by the query plan */ 3705 assert( wctrlFlags & WHERE_ORDERBY_LIMIT ); 3706 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 3707 if( j>=pLoop->nLTerm ) continue; 3708 } 3709 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 3710 if( sqlite3ExprCollSeqMatch(pWInfo->pParse, 3711 pOrderBy->a[i].pExpr, pTerm->pExpr)==0 ){ 3712 continue; 3713 } 3714 testcase( pTerm->pExpr->op==TK_IS ); 3715 } 3716 obSat |= MASKBIT(i); 3717 } 3718 3719 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 3720 if( pLoop->wsFlags & WHERE_IPK ){ 3721 pIndex = 0; 3722 nKeyCol = 0; 3723 nColumn = 1; 3724 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 3725 return 0; 3726 }else{ 3727 nKeyCol = pIndex->nKeyCol; 3728 nColumn = pIndex->nColumn; 3729 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 3730 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 3731 || !HasRowid(pIndex->pTable)); 3732 isOrderDistinct = IsUniqueIndex(pIndex); 3733 } 3734 3735 /* Loop through all columns of the index and deal with the ones 3736 ** that are not constrained by == or IN. 3737 */ 3738 rev = revSet = 0; 3739 distinctColumns = 0; 3740 for(j=0; j<nColumn; j++){ 3741 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 3742 3743 assert( j>=pLoop->u.btree.nEq 3744 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 3745 ); 3746 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 3747 u16 eOp = pLoop->aLTerm[j]->eOperator; 3748 3749 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 3750 ** doing WHERE_ORDERBY_LIMIT processing). 3751 ** 3752 ** If the current term is a column of an ((?,?) IN (SELECT...)) 3753 ** expression for which the SELECT returns more than one column, 3754 ** check that it is the only column used by this loop. Otherwise, 3755 ** if it is one of two or more, none of the columns can be 3756 ** considered to match an ORDER BY term. */ 3757 if( (eOp & eqOpMask)!=0 ){ 3758 if( eOp & WO_ISNULL ){ 3759 testcase( isOrderDistinct ); 3760 isOrderDistinct = 0; 3761 } 3762 continue; 3763 }else if( ALWAYS(eOp & WO_IN) ){ 3764 /* ALWAYS() justification: eOp is an equality operator due to the 3765 ** j<pLoop->u.btree.nEq constraint above. Any equality other 3766 ** than WO_IN is captured by the previous "if". So this one 3767 ** always has to be WO_IN. */ 3768 Expr *pX = pLoop->aLTerm[j]->pExpr; 3769 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 3770 if( pLoop->aLTerm[i]->pExpr==pX ){ 3771 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 3772 bOnce = 0; 3773 break; 3774 } 3775 } 3776 } 3777 } 3778 3779 /* Get the column number in the table (iColumn) and sort order 3780 ** (revIdx) for the j-th column of the index. 3781 */ 3782 if( pIndex ){ 3783 iColumn = pIndex->aiColumn[j]; 3784 revIdx = pIndex->aSortOrder[j]; 3785 if( iColumn==pIndex->pTable->iPKey ) iColumn = XN_ROWID; 3786 }else{ 3787 iColumn = XN_ROWID; 3788 revIdx = 0; 3789 } 3790 3791 /* An unconstrained column that might be NULL means that this 3792 ** WhereLoop is not well-ordered 3793 */ 3794 if( isOrderDistinct 3795 && iColumn>=0 3796 && j>=pLoop->u.btree.nEq 3797 && pIndex->pTable->aCol[iColumn].notNull==0 3798 ){ 3799 isOrderDistinct = 0; 3800 } 3801 3802 /* Find the ORDER BY term that corresponds to the j-th column 3803 ** of the index and mark that ORDER BY term off 3804 */ 3805 isMatch = 0; 3806 for(i=0; bOnce && i<nOrderBy; i++){ 3807 if( MASKBIT(i) & obSat ) continue; 3808 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr); 3809 testcase( wctrlFlags & WHERE_GROUPBY ); 3810 testcase( wctrlFlags & WHERE_DISTINCTBY ); 3811 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 3812 if( iColumn>=XN_ROWID ){ 3813 if( pOBExpr->op!=TK_COLUMN ) continue; 3814 if( pOBExpr->iTable!=iCur ) continue; 3815 if( pOBExpr->iColumn!=iColumn ) continue; 3816 }else{ 3817 Expr *pIdxExpr = pIndex->aColExpr->a[j].pExpr; 3818 if( sqlite3ExprCompareSkip(pOBExpr, pIdxExpr, iCur) ){ 3819 continue; 3820 } 3821 } 3822 if( iColumn!=XN_ROWID ){ 3823 pColl = sqlite3ExprNNCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3824 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 3825 } 3826 pLoop->u.btree.nIdxCol = j+1; 3827 isMatch = 1; 3828 break; 3829 } 3830 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 3831 /* Make sure the sort order is compatible in an ORDER BY clause. 3832 ** Sort order is irrelevant for a GROUP BY clause. */ 3833 if( revSet ){ 3834 if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0; 3835 }else{ 3836 rev = revIdx ^ pOrderBy->a[i].sortOrder; 3837 if( rev ) *pRevMask |= MASKBIT(iLoop); 3838 revSet = 1; 3839 } 3840 } 3841 if( isMatch ){ 3842 if( iColumn==XN_ROWID ){ 3843 testcase( distinctColumns==0 ); 3844 distinctColumns = 1; 3845 } 3846 obSat |= MASKBIT(i); 3847 }else{ 3848 /* No match found */ 3849 if( j==0 || j<nKeyCol ){ 3850 testcase( isOrderDistinct!=0 ); 3851 isOrderDistinct = 0; 3852 } 3853 break; 3854 } 3855 } /* end Loop over all index columns */ 3856 if( distinctColumns ){ 3857 testcase( isOrderDistinct==0 ); 3858 isOrderDistinct = 1; 3859 } 3860 } /* end-if not one-row */ 3861 3862 /* Mark off any other ORDER BY terms that reference pLoop */ 3863 if( isOrderDistinct ){ 3864 orderDistinctMask |= pLoop->maskSelf; 3865 for(i=0; i<nOrderBy; i++){ 3866 Expr *p; 3867 Bitmask mTerm; 3868 if( MASKBIT(i) & obSat ) continue; 3869 p = pOrderBy->a[i].pExpr; 3870 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 3871 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 3872 if( (mTerm&~orderDistinctMask)==0 ){ 3873 obSat |= MASKBIT(i); 3874 } 3875 } 3876 } 3877 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 3878 if( obSat==obDone ) return (i8)nOrderBy; 3879 if( !isOrderDistinct ){ 3880 for(i=nOrderBy-1; i>0; i--){ 3881 Bitmask m = MASKBIT(i) - 1; 3882 if( (obSat&m)==m ) return i; 3883 } 3884 return 0; 3885 } 3886 return -1; 3887 } 3888 3889 3890 /* 3891 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 3892 ** the planner assumes that the specified pOrderBy list is actually a GROUP 3893 ** BY clause - and so any order that groups rows as required satisfies the 3894 ** request. 3895 ** 3896 ** Normally, in this case it is not possible for the caller to determine 3897 ** whether or not the rows are really being delivered in sorted order, or 3898 ** just in some other order that provides the required grouping. However, 3899 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 3900 ** this function may be called on the returned WhereInfo object. It returns 3901 ** true if the rows really will be sorted in the specified order, or false 3902 ** otherwise. 3903 ** 3904 ** For example, assuming: 3905 ** 3906 ** CREATE INDEX i1 ON t1(x, Y); 3907 ** 3908 ** then 3909 ** 3910 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 3911 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 3912 */ 3913 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 3914 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 3915 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 3916 return pWInfo->sorted; 3917 } 3918 3919 #ifdef WHERETRACE_ENABLED 3920 /* For debugging use only: */ 3921 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 3922 static char zName[65]; 3923 int i; 3924 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 3925 if( pLast ) zName[i++] = pLast->cId; 3926 zName[i] = 0; 3927 return zName; 3928 } 3929 #endif 3930 3931 /* 3932 ** Return the cost of sorting nRow rows, assuming that the keys have 3933 ** nOrderby columns and that the first nSorted columns are already in 3934 ** order. 3935 */ 3936 static LogEst whereSortingCost( 3937 WhereInfo *pWInfo, 3938 LogEst nRow, 3939 int nOrderBy, 3940 int nSorted 3941 ){ 3942 /* TUNING: Estimated cost of a full external sort, where N is 3943 ** the number of rows to sort is: 3944 ** 3945 ** cost = (3.0 * N * log(N)). 3946 ** 3947 ** Or, if the order-by clause has X terms but only the last Y 3948 ** terms are out of order, then block-sorting will reduce the 3949 ** sorting cost to: 3950 ** 3951 ** cost = (3.0 * N * log(N)) * (Y/X) 3952 ** 3953 ** The (Y/X) term is implemented using stack variable rScale 3954 ** below. */ 3955 LogEst rScale, rSortCost; 3956 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 3957 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 3958 rSortCost = nRow + rScale + 16; 3959 3960 /* Multiple by log(M) where M is the number of output rows. 3961 ** Use the LIMIT for M if it is smaller */ 3962 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 3963 nRow = pWInfo->iLimit; 3964 } 3965 rSortCost += estLog(nRow); 3966 return rSortCost; 3967 } 3968 3969 /* 3970 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 3971 ** attempts to find the lowest cost path that visits each WhereLoop 3972 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 3973 ** 3974 ** Assume that the total number of output rows that will need to be sorted 3975 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 3976 ** costs if nRowEst==0. 3977 ** 3978 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 3979 ** error occurs. 3980 */ 3981 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 3982 int mxChoice; /* Maximum number of simultaneous paths tracked */ 3983 int nLoop; /* Number of terms in the join */ 3984 Parse *pParse; /* Parsing context */ 3985 sqlite3 *db; /* The database connection */ 3986 int iLoop; /* Loop counter over the terms of the join */ 3987 int ii, jj; /* Loop counters */ 3988 int mxI = 0; /* Index of next entry to replace */ 3989 int nOrderBy; /* Number of ORDER BY clause terms */ 3990 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 3991 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 3992 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 3993 WherePath *aFrom; /* All nFrom paths at the previous level */ 3994 WherePath *aTo; /* The nTo best paths at the current level */ 3995 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 3996 WherePath *pTo; /* An element of aTo[] that we are working on */ 3997 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 3998 WhereLoop **pX; /* Used to divy up the pSpace memory */ 3999 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 4000 char *pSpace; /* Temporary memory used by this routine */ 4001 int nSpace; /* Bytes of space allocated at pSpace */ 4002 4003 pParse = pWInfo->pParse; 4004 db = pParse->db; 4005 nLoop = pWInfo->nLevel; 4006 /* TUNING: For simple queries, only the best path is tracked. 4007 ** For 2-way joins, the 5 best paths are followed. 4008 ** For joins of 3 or more tables, track the 10 best paths */ 4009 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 4010 assert( nLoop<=pWInfo->pTabList->nSrc ); 4011 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 4012 4013 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 4014 ** case the purpose of this call is to estimate the number of rows returned 4015 ** by the overall query. Once this estimate has been obtained, the caller 4016 ** will invoke this function a second time, passing the estimate as the 4017 ** nRowEst parameter. */ 4018 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 4019 nOrderBy = 0; 4020 }else{ 4021 nOrderBy = pWInfo->pOrderBy->nExpr; 4022 } 4023 4024 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 4025 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 4026 nSpace += sizeof(LogEst) * nOrderBy; 4027 pSpace = sqlite3DbMallocRawNN(db, nSpace); 4028 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 4029 aTo = (WherePath*)pSpace; 4030 aFrom = aTo+mxChoice; 4031 memset(aFrom, 0, sizeof(aFrom[0])); 4032 pX = (WhereLoop**)(aFrom+mxChoice); 4033 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 4034 pFrom->aLoop = pX; 4035 } 4036 if( nOrderBy ){ 4037 /* If there is an ORDER BY clause and it is not being ignored, set up 4038 ** space for the aSortCost[] array. Each element of the aSortCost array 4039 ** is either zero - meaning it has not yet been initialized - or the 4040 ** cost of sorting nRowEst rows of data where the first X terms of 4041 ** the ORDER BY clause are already in order, where X is the array 4042 ** index. */ 4043 aSortCost = (LogEst*)pX; 4044 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 4045 } 4046 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 4047 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 4048 4049 /* Seed the search with a single WherePath containing zero WhereLoops. 4050 ** 4051 ** TUNING: Do not let the number of iterations go above 28. If the cost 4052 ** of computing an automatic index is not paid back within the first 28 4053 ** rows, then do not use the automatic index. */ 4054 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 4055 nFrom = 1; 4056 assert( aFrom[0].isOrdered==0 ); 4057 if( nOrderBy ){ 4058 /* If nLoop is zero, then there are no FROM terms in the query. Since 4059 ** in this case the query may return a maximum of one row, the results 4060 ** are already in the requested order. Set isOrdered to nOrderBy to 4061 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 4062 ** -1, indicating that the result set may or may not be ordered, 4063 ** depending on the loops added to the current plan. */ 4064 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 4065 } 4066 4067 /* Compute successively longer WherePaths using the previous generation 4068 ** of WherePaths as the basis for the next. Keep track of the mxChoice 4069 ** best paths at each generation */ 4070 for(iLoop=0; iLoop<nLoop; iLoop++){ 4071 nTo = 0; 4072 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 4073 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 4074 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 4075 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 4076 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 4077 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 4078 Bitmask maskNew; /* Mask of src visited by (..) */ 4079 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 4080 4081 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 4082 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 4083 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<3 ){ 4084 /* Do not use an automatic index if the this loop is expected 4085 ** to run less than 1.25 times. It is tempting to also exclude 4086 ** automatic index usage on an outer loop, but sometimes an automatic 4087 ** index is useful in the outer loop of a correlated subquery. */ 4088 assert( 10==sqlite3LogEst(2) ); 4089 continue; 4090 } 4091 4092 /* At this point, pWLoop is a candidate to be the next loop. 4093 ** Compute its cost */ 4094 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 4095 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 4096 nOut = pFrom->nRow + pWLoop->nOut; 4097 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 4098 if( isOrdered<0 ){ 4099 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 4100 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 4101 iLoop, pWLoop, &revMask); 4102 }else{ 4103 revMask = pFrom->revLoop; 4104 } 4105 if( isOrdered>=0 && isOrdered<nOrderBy ){ 4106 if( aSortCost[isOrdered]==0 ){ 4107 aSortCost[isOrdered] = whereSortingCost( 4108 pWInfo, nRowEst, nOrderBy, isOrdered 4109 ); 4110 } 4111 /* TUNING: Add a small extra penalty (5) to sorting as an 4112 ** extra encouragment to the query planner to select a plan 4113 ** where the rows emerge in the correct order without any sorting 4114 ** required. */ 4115 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]) + 5; 4116 4117 WHERETRACE(0x002, 4118 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 4119 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 4120 rUnsorted, rCost)); 4121 }else{ 4122 rCost = rUnsorted; 4123 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 4124 } 4125 4126 /* Check to see if pWLoop should be added to the set of 4127 ** mxChoice best-so-far paths. 4128 ** 4129 ** First look for an existing path among best-so-far paths 4130 ** that covers the same set of loops and has the same isOrdered 4131 ** setting as the current path candidate. 4132 ** 4133 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 4134 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 4135 ** of legal values for isOrdered, -1..64. 4136 */ 4137 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4138 if( pTo->maskLoop==maskNew 4139 && ((pTo->isOrdered^isOrdered)&0x80)==0 4140 ){ 4141 testcase( jj==nTo-1 ); 4142 break; 4143 } 4144 } 4145 if( jj>=nTo ){ 4146 /* None of the existing best-so-far paths match the candidate. */ 4147 if( nTo>=mxChoice 4148 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4149 ){ 4150 /* The current candidate is no better than any of the mxChoice 4151 ** paths currently in the best-so-far buffer. So discard 4152 ** this candidate as not viable. */ 4153 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4154 if( sqlite3WhereTrace&0x4 ){ 4155 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4156 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4157 isOrdered>=0 ? isOrdered+'0' : '?'); 4158 } 4159 #endif 4160 continue; 4161 } 4162 /* If we reach this points it means that the new candidate path 4163 ** needs to be added to the set of best-so-far paths. */ 4164 if( nTo<mxChoice ){ 4165 /* Increase the size of the aTo set by one */ 4166 jj = nTo++; 4167 }else{ 4168 /* New path replaces the prior worst to keep count below mxChoice */ 4169 jj = mxI; 4170 } 4171 pTo = &aTo[jj]; 4172 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4173 if( sqlite3WhereTrace&0x4 ){ 4174 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4175 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4176 isOrdered>=0 ? isOrdered+'0' : '?'); 4177 } 4178 #endif 4179 }else{ 4180 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4181 ** same set of loops and has the same isOrdered setting as the 4182 ** candidate path. Check to see if the candidate should replace 4183 ** pTo or if the candidate should be skipped. 4184 ** 4185 ** The conditional is an expanded vector comparison equivalent to: 4186 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4187 */ 4188 if( pTo->rCost<rCost 4189 || (pTo->rCost==rCost 4190 && (pTo->nRow<nOut 4191 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4192 ) 4193 ) 4194 ){ 4195 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4196 if( sqlite3WhereTrace&0x4 ){ 4197 sqlite3DebugPrintf( 4198 "Skip %s cost=%-3d,%3d,%3d order=%c", 4199 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4200 isOrdered>=0 ? isOrdered+'0' : '?'); 4201 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4202 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4203 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4204 } 4205 #endif 4206 /* Discard the candidate path from further consideration */ 4207 testcase( pTo->rCost==rCost ); 4208 continue; 4209 } 4210 testcase( pTo->rCost==rCost+1 ); 4211 /* Control reaches here if the candidate path is better than the 4212 ** pTo path. Replace pTo with the candidate. */ 4213 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4214 if( sqlite3WhereTrace&0x4 ){ 4215 sqlite3DebugPrintf( 4216 "Update %s cost=%-3d,%3d,%3d order=%c", 4217 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4218 isOrdered>=0 ? isOrdered+'0' : '?'); 4219 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4220 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4221 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4222 } 4223 #endif 4224 } 4225 /* pWLoop is a winner. Add it to the set of best so far */ 4226 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4227 pTo->revLoop = revMask; 4228 pTo->nRow = nOut; 4229 pTo->rCost = rCost; 4230 pTo->rUnsorted = rUnsorted; 4231 pTo->isOrdered = isOrdered; 4232 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4233 pTo->aLoop[iLoop] = pWLoop; 4234 if( nTo>=mxChoice ){ 4235 mxI = 0; 4236 mxCost = aTo[0].rCost; 4237 mxUnsorted = aTo[0].nRow; 4238 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4239 if( pTo->rCost>mxCost 4240 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4241 ){ 4242 mxCost = pTo->rCost; 4243 mxUnsorted = pTo->rUnsorted; 4244 mxI = jj; 4245 } 4246 } 4247 } 4248 } 4249 } 4250 4251 #ifdef WHERETRACE_ENABLED /* >=2 */ 4252 if( sqlite3WhereTrace & 0x02 ){ 4253 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4254 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4255 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4256 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4257 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4258 if( pTo->isOrdered>0 ){ 4259 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4260 }else{ 4261 sqlite3DebugPrintf("\n"); 4262 } 4263 } 4264 } 4265 #endif 4266 4267 /* Swap the roles of aFrom and aTo for the next generation */ 4268 pFrom = aTo; 4269 aTo = aFrom; 4270 aFrom = pFrom; 4271 nFrom = nTo; 4272 } 4273 4274 if( nFrom==0 ){ 4275 sqlite3ErrorMsg(pParse, "no query solution"); 4276 sqlite3DbFreeNN(db, pSpace); 4277 return SQLITE_ERROR; 4278 } 4279 4280 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4281 pFrom = aFrom; 4282 for(ii=1; ii<nFrom; ii++){ 4283 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4284 } 4285 assert( pWInfo->nLevel==nLoop ); 4286 /* Load the lowest cost path into pWInfo */ 4287 for(iLoop=0; iLoop<nLoop; iLoop++){ 4288 WhereLevel *pLevel = pWInfo->a + iLoop; 4289 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4290 pLevel->iFrom = pWLoop->iTab; 4291 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4292 } 4293 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4294 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4295 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4296 && nRowEst 4297 ){ 4298 Bitmask notUsed; 4299 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4300 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4301 if( rc==pWInfo->pResultSet->nExpr ){ 4302 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4303 } 4304 } 4305 pWInfo->bOrderedInnerLoop = 0; 4306 if( pWInfo->pOrderBy ){ 4307 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4308 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4309 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4310 } 4311 }else{ 4312 pWInfo->nOBSat = pFrom->isOrdered; 4313 pWInfo->revMask = pFrom->revLoop; 4314 if( pWInfo->nOBSat<=0 ){ 4315 pWInfo->nOBSat = 0; 4316 if( nLoop>0 ){ 4317 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4318 if( (wsFlags & WHERE_ONEROW)==0 4319 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4320 ){ 4321 Bitmask m = 0; 4322 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4323 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4324 testcase( wsFlags & WHERE_IPK ); 4325 testcase( wsFlags & WHERE_COLUMN_IN ); 4326 if( rc==pWInfo->pOrderBy->nExpr ){ 4327 pWInfo->bOrderedInnerLoop = 1; 4328 pWInfo->revMask = m; 4329 } 4330 } 4331 } 4332 } 4333 } 4334 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4335 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4336 ){ 4337 Bitmask revMask = 0; 4338 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4339 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4340 ); 4341 assert( pWInfo->sorted==0 ); 4342 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4343 pWInfo->sorted = 1; 4344 pWInfo->revMask = revMask; 4345 } 4346 } 4347 } 4348 4349 4350 pWInfo->nRowOut = pFrom->nRow; 4351 4352 /* Free temporary memory and return success */ 4353 sqlite3DbFreeNN(db, pSpace); 4354 return SQLITE_OK; 4355 } 4356 4357 /* 4358 ** Most queries use only a single table (they are not joins) and have 4359 ** simple == constraints against indexed fields. This routine attempts 4360 ** to plan those simple cases using much less ceremony than the 4361 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4362 ** times for the common case. 4363 ** 4364 ** Return non-zero on success, if this query can be handled by this 4365 ** no-frills query planner. Return zero if this query needs the 4366 ** general-purpose query planner. 4367 */ 4368 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4369 WhereInfo *pWInfo; 4370 struct SrcList_item *pItem; 4371 WhereClause *pWC; 4372 WhereTerm *pTerm; 4373 WhereLoop *pLoop; 4374 int iCur; 4375 int j; 4376 Table *pTab; 4377 Index *pIdx; 4378 4379 pWInfo = pBuilder->pWInfo; 4380 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4381 assert( pWInfo->pTabList->nSrc>=1 ); 4382 pItem = pWInfo->pTabList->a; 4383 pTab = pItem->pTab; 4384 if( IsVirtual(pTab) ) return 0; 4385 if( pItem->fg.isIndexedBy ) return 0; 4386 iCur = pItem->iCursor; 4387 pWC = &pWInfo->sWC; 4388 pLoop = pBuilder->pNew; 4389 pLoop->wsFlags = 0; 4390 pLoop->nSkip = 0; 4391 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0); 4392 if( pTerm ){ 4393 testcase( pTerm->eOperator & WO_IS ); 4394 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4395 pLoop->aLTerm[0] = pTerm; 4396 pLoop->nLTerm = 1; 4397 pLoop->u.btree.nEq = 1; 4398 /* TUNING: Cost of a rowid lookup is 10 */ 4399 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 4400 }else{ 4401 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 4402 int opMask; 4403 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 4404 if( !IsUniqueIndex(pIdx) 4405 || pIdx->pPartIdxWhere!=0 4406 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 4407 ) continue; 4408 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 4409 for(j=0; j<pIdx->nKeyCol; j++){ 4410 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx); 4411 if( pTerm==0 ) break; 4412 testcase( pTerm->eOperator & WO_IS ); 4413 pLoop->aLTerm[j] = pTerm; 4414 } 4415 if( j!=pIdx->nKeyCol ) continue; 4416 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 4417 if( pIdx->isCovering || (pItem->colUsed & pIdx->colNotIdxed)==0 ){ 4418 pLoop->wsFlags |= WHERE_IDX_ONLY; 4419 } 4420 pLoop->nLTerm = j; 4421 pLoop->u.btree.nEq = j; 4422 pLoop->u.btree.pIndex = pIdx; 4423 /* TUNING: Cost of a unique index lookup is 15 */ 4424 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 4425 break; 4426 } 4427 } 4428 if( pLoop->wsFlags ){ 4429 pLoop->nOut = (LogEst)1; 4430 pWInfo->a[0].pWLoop = pLoop; 4431 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 4432 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 4433 pWInfo->a[0].iTabCur = iCur; 4434 pWInfo->nRowOut = 1; 4435 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 4436 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 4437 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4438 } 4439 #ifdef SQLITE_DEBUG 4440 pLoop->cId = '0'; 4441 #endif 4442 return 1; 4443 } 4444 return 0; 4445 } 4446 4447 /* 4448 ** Helper function for exprIsDeterministic(). 4449 */ 4450 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 4451 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 4452 pWalker->eCode = 0; 4453 return WRC_Abort; 4454 } 4455 return WRC_Continue; 4456 } 4457 4458 /* 4459 ** Return true if the expression contains no non-deterministic SQL 4460 ** functions. Do not consider non-deterministic SQL functions that are 4461 ** part of sub-select statements. 4462 */ 4463 static int exprIsDeterministic(Expr *p){ 4464 Walker w; 4465 memset(&w, 0, sizeof(w)); 4466 w.eCode = 1; 4467 w.xExprCallback = exprNodeIsDeterministic; 4468 w.xSelectCallback = sqlite3SelectWalkFail; 4469 sqlite3WalkExpr(&w, p); 4470 return w.eCode; 4471 } 4472 4473 /* 4474 ** Generate the beginning of the loop used for WHERE clause processing. 4475 ** The return value is a pointer to an opaque structure that contains 4476 ** information needed to terminate the loop. Later, the calling routine 4477 ** should invoke sqlite3WhereEnd() with the return value of this function 4478 ** in order to complete the WHERE clause processing. 4479 ** 4480 ** If an error occurs, this routine returns NULL. 4481 ** 4482 ** The basic idea is to do a nested loop, one loop for each table in 4483 ** the FROM clause of a select. (INSERT and UPDATE statements are the 4484 ** same as a SELECT with only a single table in the FROM clause.) For 4485 ** example, if the SQL is this: 4486 ** 4487 ** SELECT * FROM t1, t2, t3 WHERE ...; 4488 ** 4489 ** Then the code generated is conceptually like the following: 4490 ** 4491 ** foreach row1 in t1 do \ Code generated 4492 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 4493 ** foreach row3 in t3 do / 4494 ** ... 4495 ** end \ Code generated 4496 ** end |-- by sqlite3WhereEnd() 4497 ** end / 4498 ** 4499 ** Note that the loops might not be nested in the order in which they 4500 ** appear in the FROM clause if a different order is better able to make 4501 ** use of indices. Note also that when the IN operator appears in 4502 ** the WHERE clause, it might result in additional nested loops for 4503 ** scanning through all values on the right-hand side of the IN. 4504 ** 4505 ** There are Btree cursors associated with each table. t1 uses cursor 4506 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 4507 ** And so forth. This routine generates code to open those VDBE cursors 4508 ** and sqlite3WhereEnd() generates the code to close them. 4509 ** 4510 ** The code that sqlite3WhereBegin() generates leaves the cursors named 4511 ** in pTabList pointing at their appropriate entries. The [...] code 4512 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 4513 ** data from the various tables of the loop. 4514 ** 4515 ** If the WHERE clause is empty, the foreach loops must each scan their 4516 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 4517 ** the tables have indices and there are terms in the WHERE clause that 4518 ** refer to those indices, a complete table scan can be avoided and the 4519 ** code will run much faster. Most of the work of this routine is checking 4520 ** to see if there are indices that can be used to speed up the loop. 4521 ** 4522 ** Terms of the WHERE clause are also used to limit which rows actually 4523 ** make it to the "..." in the middle of the loop. After each "foreach", 4524 ** terms of the WHERE clause that use only terms in that loop and outer 4525 ** loops are evaluated and if false a jump is made around all subsequent 4526 ** inner loops (or around the "..." if the test occurs within the inner- 4527 ** most loop) 4528 ** 4529 ** OUTER JOINS 4530 ** 4531 ** An outer join of tables t1 and t2 is conceptally coded as follows: 4532 ** 4533 ** foreach row1 in t1 do 4534 ** flag = 0 4535 ** foreach row2 in t2 do 4536 ** start: 4537 ** ... 4538 ** flag = 1 4539 ** end 4540 ** if flag==0 then 4541 ** move the row2 cursor to a null row 4542 ** goto start 4543 ** fi 4544 ** end 4545 ** 4546 ** ORDER BY CLAUSE PROCESSING 4547 ** 4548 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 4549 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 4550 ** if there is one. If there is no ORDER BY clause or if this routine 4551 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 4552 ** 4553 ** The iIdxCur parameter is the cursor number of an index. If 4554 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 4555 ** to use for OR clause processing. The WHERE clause should use this 4556 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 4557 ** the first cursor in an array of cursors for all indices. iIdxCur should 4558 ** be used to compute the appropriate cursor depending on which index is 4559 ** used. 4560 */ 4561 WhereInfo *sqlite3WhereBegin( 4562 Parse *pParse, /* The parser context */ 4563 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 4564 Expr *pWhere, /* The WHERE clause */ 4565 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 4566 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 4567 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 4568 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 4569 ** If WHERE_USE_LIMIT, then the limit amount */ 4570 ){ 4571 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4572 int nTabList; /* Number of elements in pTabList */ 4573 WhereInfo *pWInfo; /* Will become the return value of this function */ 4574 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4575 Bitmask notReady; /* Cursors that are not yet positioned */ 4576 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 4577 WhereMaskSet *pMaskSet; /* The expression mask set */ 4578 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 4579 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 4580 int ii; /* Loop counter */ 4581 sqlite3 *db; /* Database connection */ 4582 int rc; /* Return code */ 4583 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 4584 4585 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 4586 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 4587 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4588 )); 4589 4590 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 4591 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4592 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 4593 4594 /* Variable initialization */ 4595 db = pParse->db; 4596 memset(&sWLB, 0, sizeof(sWLB)); 4597 4598 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 4599 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 4600 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 4601 sWLB.pOrderBy = pOrderBy; 4602 4603 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 4604 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 4605 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 4606 wctrlFlags &= ~WHERE_WANT_DISTINCT; 4607 } 4608 4609 /* The number of tables in the FROM clause is limited by the number of 4610 ** bits in a Bitmask 4611 */ 4612 testcase( pTabList->nSrc==BMS ); 4613 if( pTabList->nSrc>BMS ){ 4614 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 4615 return 0; 4616 } 4617 4618 /* This function normally generates a nested loop for all tables in 4619 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 4620 ** only generate code for the first table in pTabList and assume that 4621 ** any cursors associated with subsequent tables are uninitialized. 4622 */ 4623 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 4624 4625 /* Allocate and initialize the WhereInfo structure that will become the 4626 ** return value. A single allocation is used to store the WhereInfo 4627 ** struct, the contents of WhereInfo.a[], the WhereClause structure 4628 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 4629 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 4630 ** some architectures. Hence the ROUND8() below. 4631 */ 4632 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 4633 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 4634 if( db->mallocFailed ){ 4635 sqlite3DbFree(db, pWInfo); 4636 pWInfo = 0; 4637 goto whereBeginError; 4638 } 4639 pWInfo->pParse = pParse; 4640 pWInfo->pTabList = pTabList; 4641 pWInfo->pOrderBy = pOrderBy; 4642 pWInfo->pWhere = pWhere; 4643 pWInfo->pResultSet = pResultSet; 4644 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 4645 pWInfo->nLevel = nTabList; 4646 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(v); 4647 pWInfo->wctrlFlags = wctrlFlags; 4648 pWInfo->iLimit = iAuxArg; 4649 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 4650 memset(&pWInfo->nOBSat, 0, 4651 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 4652 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 4653 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 4654 pMaskSet = &pWInfo->sMaskSet; 4655 sWLB.pWInfo = pWInfo; 4656 sWLB.pWC = &pWInfo->sWC; 4657 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 4658 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 4659 whereLoopInit(sWLB.pNew); 4660 #ifdef SQLITE_DEBUG 4661 sWLB.pNew->cId = '*'; 4662 #endif 4663 4664 /* Split the WHERE clause into separate subexpressions where each 4665 ** subexpression is separated by an AND operator. 4666 */ 4667 initMaskSet(pMaskSet); 4668 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 4669 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 4670 4671 /* Special case: No FROM clause 4672 */ 4673 if( nTabList==0 ){ 4674 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 4675 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4676 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4677 } 4678 ExplainQueryPlan((pParse, 0, "SCAN CONSTANT ROW")); 4679 }else{ 4680 /* Assign a bit from the bitmask to every term in the FROM clause. 4681 ** 4682 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 4683 ** 4684 ** The rule of the previous sentence ensures thta if X is the bitmask for 4685 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 4686 ** Knowing the bitmask for all tables to the left of a left join is 4687 ** important. Ticket #3015. 4688 ** 4689 ** Note that bitmasks are created for all pTabList->nSrc tables in 4690 ** pTabList, not just the first nTabList tables. nTabList is normally 4691 ** equal to pTabList->nSrc but might be shortened to 1 if the 4692 ** WHERE_OR_SUBCLAUSE flag is set. 4693 */ 4694 ii = 0; 4695 do{ 4696 createMask(pMaskSet, pTabList->a[ii].iCursor); 4697 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 4698 }while( (++ii)<pTabList->nSrc ); 4699 #ifdef SQLITE_DEBUG 4700 { 4701 Bitmask mx = 0; 4702 for(ii=0; ii<pTabList->nSrc; ii++){ 4703 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 4704 assert( m>=mx ); 4705 mx = m; 4706 } 4707 } 4708 #endif 4709 } 4710 4711 /* Analyze all of the subexpressions. */ 4712 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 4713 if( db->mallocFailed ) goto whereBeginError; 4714 4715 /* Special case: WHERE terms that do not refer to any tables in the join 4716 ** (constant expressions). Evaluate each such term, and jump over all the 4717 ** generated code if the result is not true. 4718 ** 4719 ** Do not do this if the expression contains non-deterministic functions 4720 ** that are not within a sub-select. This is not strictly required, but 4721 ** preserves SQLite's legacy behaviour in the following two cases: 4722 ** 4723 ** FROM ... WHERE random()>0; -- eval random() once per row 4724 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 4725 */ 4726 for(ii=0; ii<sWLB.pWC->nTerm; ii++){ 4727 WhereTerm *pT = &sWLB.pWC->a[ii]; 4728 if( pT->wtFlags & TERM_VIRTUAL ) continue; 4729 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 4730 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 4731 pT->wtFlags |= TERM_CODED; 4732 } 4733 } 4734 4735 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4736 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 4737 /* The DISTINCT marking is pointless. Ignore it. */ 4738 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4739 }else if( pOrderBy==0 ){ 4740 /* Try to ORDER BY the result set to make distinct processing easier */ 4741 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 4742 pWInfo->pOrderBy = pResultSet; 4743 } 4744 } 4745 4746 /* Construct the WhereLoop objects */ 4747 #if defined(WHERETRACE_ENABLED) 4748 if( sqlite3WhereTrace & 0xffff ){ 4749 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 4750 if( wctrlFlags & WHERE_USE_LIMIT ){ 4751 sqlite3DebugPrintf(", limit: %d", iAuxArg); 4752 } 4753 sqlite3DebugPrintf(")\n"); 4754 } 4755 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 4756 sqlite3WhereClausePrint(sWLB.pWC); 4757 } 4758 #endif 4759 4760 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 4761 rc = whereLoopAddAll(&sWLB); 4762 if( rc ) goto whereBeginError; 4763 4764 #ifdef WHERETRACE_ENABLED 4765 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4766 WhereLoop *p; 4767 int i; 4768 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4769 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4770 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4771 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 4772 whereLoopPrint(p, sWLB.pWC); 4773 } 4774 } 4775 #endif 4776 4777 wherePathSolver(pWInfo, 0); 4778 if( db->mallocFailed ) goto whereBeginError; 4779 if( pWInfo->pOrderBy ){ 4780 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 4781 if( db->mallocFailed ) goto whereBeginError; 4782 } 4783 } 4784 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 4785 pWInfo->revMask = ALLBITS; 4786 } 4787 if( pParse->nErr || NEVER(db->mallocFailed) ){ 4788 goto whereBeginError; 4789 } 4790 #ifdef WHERETRACE_ENABLED 4791 if( sqlite3WhereTrace ){ 4792 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 4793 if( pWInfo->nOBSat>0 ){ 4794 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 4795 } 4796 switch( pWInfo->eDistinct ){ 4797 case WHERE_DISTINCT_UNIQUE: { 4798 sqlite3DebugPrintf(" DISTINCT=unique"); 4799 break; 4800 } 4801 case WHERE_DISTINCT_ORDERED: { 4802 sqlite3DebugPrintf(" DISTINCT=ordered"); 4803 break; 4804 } 4805 case WHERE_DISTINCT_UNORDERED: { 4806 sqlite3DebugPrintf(" DISTINCT=unordered"); 4807 break; 4808 } 4809 } 4810 sqlite3DebugPrintf("\n"); 4811 for(ii=0; ii<pWInfo->nLevel; ii++){ 4812 whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 4813 } 4814 } 4815 #endif 4816 4817 /* Attempt to omit tables from the join that do not affect the result. 4818 ** For a table to not affect the result, the following must be true: 4819 ** 4820 ** 1) The query must not be an aggregate. 4821 ** 2) The table must be the RHS of a LEFT JOIN. 4822 ** 3) Either the query must be DISTINCT, or else the ON or USING clause 4823 ** must contain a constraint that limits the scan of the table to 4824 ** at most a single row. 4825 ** 4) The table must not be referenced by any part of the query apart 4826 ** from its own USING or ON clause. 4827 ** 4828 ** For example, given: 4829 ** 4830 ** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1); 4831 ** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2); 4832 ** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3); 4833 ** 4834 ** then table t2 can be omitted from the following: 4835 ** 4836 ** SELECT v1, v3 FROM t1 4837 ** LEFT JOIN t2 USING (t1.ipk=t2.ipk) 4838 ** LEFT JOIN t3 USING (t1.ipk=t3.ipk) 4839 ** 4840 ** or from: 4841 ** 4842 ** SELECT DISTINCT v1, v3 FROM t1 4843 ** LEFT JOIN t2 4844 ** LEFT JOIN t3 USING (t1.ipk=t3.ipk) 4845 */ 4846 notReady = ~(Bitmask)0; 4847 if( pWInfo->nLevel>=2 4848 && pResultSet!=0 /* guarantees condition (1) above */ 4849 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 4850 ){ 4851 int i; 4852 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet); 4853 if( sWLB.pOrderBy ){ 4854 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy); 4855 } 4856 for(i=pWInfo->nLevel-1; i>=1; i--){ 4857 WhereTerm *pTerm, *pEnd; 4858 struct SrcList_item *pItem; 4859 pLoop = pWInfo->a[i].pWLoop; 4860 pItem = &pWInfo->pTabList->a[pLoop->iTab]; 4861 if( (pItem->fg.jointype & JT_LEFT)==0 ) continue; 4862 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 4863 && (pLoop->wsFlags & WHERE_ONEROW)==0 4864 ){ 4865 continue; 4866 } 4867 if( (tabUsed & pLoop->maskSelf)!=0 ) continue; 4868 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; 4869 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4870 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 4871 if( !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 4872 || pTerm->pExpr->iRightJoinTable!=pItem->iCursor 4873 ){ 4874 break; 4875 } 4876 } 4877 } 4878 if( pTerm<pEnd ) continue; 4879 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 4880 notReady &= ~pLoop->maskSelf; 4881 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 4882 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 4883 pTerm->wtFlags |= TERM_CODED; 4884 } 4885 } 4886 if( i!=pWInfo->nLevel-1 ){ 4887 int nByte = (pWInfo->nLevel-1-i) * sizeof(WhereLevel); 4888 memmove(&pWInfo->a[i], &pWInfo->a[i+1], nByte); 4889 } 4890 pWInfo->nLevel--; 4891 nTabList--; 4892 } 4893 } 4894 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 4895 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 4896 4897 /* If the caller is an UPDATE or DELETE statement that is requesting 4898 ** to use a one-pass algorithm, determine if this is appropriate. 4899 ** 4900 ** A one-pass approach can be used if the caller has requested one 4901 ** and either (a) the scan visits at most one row or (b) each 4902 ** of the following are true: 4903 ** 4904 ** * the caller has indicated that a one-pass approach can be used 4905 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and 4906 ** * the table is not a virtual table, and 4907 ** * either the scan does not use the OR optimization or the caller 4908 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified 4909 ** for DELETE). 4910 ** 4911 ** The last qualification is because an UPDATE statement uses 4912 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can 4913 ** use a one-pass approach, and this is not set accurately for scans 4914 ** that use the OR optimization. 4915 */ 4916 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 4917 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 4918 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 4919 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 4920 if( bOnerow || ( 4921 0!=(wctrlFlags & WHERE_ONEPASS_MULTIROW) 4922 && 0==(wsFlags & WHERE_VIRTUALTABLE) 4923 && (0==(wsFlags & WHERE_MULTI_OR) || (wctrlFlags & WHERE_DUPLICATES_OK)) 4924 )){ 4925 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 4926 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 4927 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 4928 bFordelete = OPFLAG_FORDELETE; 4929 } 4930 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 4931 } 4932 } 4933 } 4934 4935 /* Open all tables in the pTabList and any indices selected for 4936 ** searching those tables. 4937 */ 4938 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 4939 Table *pTab; /* Table to open */ 4940 int iDb; /* Index of database containing table/index */ 4941 struct SrcList_item *pTabItem; 4942 4943 pTabItem = &pTabList->a[pLevel->iFrom]; 4944 pTab = pTabItem->pTab; 4945 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 4946 pLoop = pLevel->pWLoop; 4947 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 4948 /* Do nothing */ 4949 }else 4950 #ifndef SQLITE_OMIT_VIRTUALTABLE 4951 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 4952 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 4953 int iCur = pTabItem->iCursor; 4954 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 4955 }else if( IsVirtual(pTab) ){ 4956 /* noop */ 4957 }else 4958 #endif 4959 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 4960 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 4961 int op = OP_OpenRead; 4962 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 4963 op = OP_OpenWrite; 4964 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 4965 }; 4966 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 4967 assert( pTabItem->iCursor==pLevel->iTabCur ); 4968 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 4969 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 4970 if( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol<BMS && HasRowid(pTab) ){ 4971 Bitmask b = pTabItem->colUsed; 4972 int n = 0; 4973 for(; b; b=b>>1, n++){} 4974 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 4975 assert( n<=pTab->nCol ); 4976 } 4977 #ifdef SQLITE_ENABLE_CURSOR_HINTS 4978 if( pLoop->u.btree.pIndex!=0 ){ 4979 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 4980 }else 4981 #endif 4982 { 4983 sqlite3VdbeChangeP5(v, bFordelete); 4984 } 4985 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 4986 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 4987 (const u8*)&pTabItem->colUsed, P4_INT64); 4988 #endif 4989 }else{ 4990 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 4991 } 4992 if( pLoop->wsFlags & WHERE_INDEXED ){ 4993 Index *pIx = pLoop->u.btree.pIndex; 4994 int iIndexCur; 4995 int op = OP_OpenRead; 4996 /* iAuxArg is always set to a positive value if ONEPASS is possible */ 4997 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 4998 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 4999 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 5000 ){ 5001 /* This is one term of an OR-optimization using the PRIMARY KEY of a 5002 ** WITHOUT ROWID table. No need for a separate index */ 5003 iIndexCur = pLevel->iTabCur; 5004 op = 0; 5005 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5006 Index *pJ = pTabItem->pTab->pIndex; 5007 iIndexCur = iAuxArg; 5008 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 5009 while( ALWAYS(pJ) && pJ!=pIx ){ 5010 iIndexCur++; 5011 pJ = pJ->pNext; 5012 } 5013 op = OP_OpenWrite; 5014 pWInfo->aiCurOnePass[1] = iIndexCur; 5015 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 5016 iIndexCur = iAuxArg; 5017 op = OP_ReopenIdx; 5018 }else{ 5019 iIndexCur = pParse->nTab++; 5020 } 5021 pLevel->iIdxCur = iIndexCur; 5022 assert( pIx->pSchema==pTab->pSchema ); 5023 assert( iIndexCur>=0 ); 5024 if( op ){ 5025 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 5026 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 5027 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 5028 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 5029 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 5030 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 5031 ){ 5032 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); /* Hint to COMDB2 */ 5033 } 5034 VdbeComment((v, "%s", pIx->zName)); 5035 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5036 { 5037 u64 colUsed = 0; 5038 int ii, jj; 5039 for(ii=0; ii<pIx->nColumn; ii++){ 5040 jj = pIx->aiColumn[ii]; 5041 if( jj<0 ) continue; 5042 if( jj>63 ) jj = 63; 5043 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 5044 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 5045 } 5046 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 5047 (u8*)&colUsed, P4_INT64); 5048 } 5049 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 5050 } 5051 } 5052 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 5053 } 5054 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 5055 if( db->mallocFailed ) goto whereBeginError; 5056 5057 /* Generate the code to do the search. Each iteration of the for 5058 ** loop below generates code for a single nested loop of the VM 5059 ** program. 5060 */ 5061 for(ii=0; ii<nTabList; ii++){ 5062 int addrExplain; 5063 int wsFlags; 5064 pLevel = &pWInfo->a[ii]; 5065 wsFlags = pLevel->pWLoop->wsFlags; 5066 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 5067 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 5068 constructAutomaticIndex(pParse, &pWInfo->sWC, 5069 &pTabList->a[pLevel->iFrom], notReady, pLevel); 5070 if( db->mallocFailed ) goto whereBeginError; 5071 } 5072 #endif 5073 addrExplain = sqlite3WhereExplainOneScan( 5074 pParse, pTabList, pLevel, wctrlFlags 5075 ); 5076 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 5077 notReady = sqlite3WhereCodeOneLoopStart(pWInfo, ii, notReady); 5078 pWInfo->iContinue = pLevel->addrCont; 5079 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 5080 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 5081 } 5082 } 5083 5084 /* Done. */ 5085 VdbeModuleComment((v, "Begin WHERE-core")); 5086 return pWInfo; 5087 5088 /* Jump here if malloc fails */ 5089 whereBeginError: 5090 if( pWInfo ){ 5091 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5092 whereInfoFree(db, pWInfo); 5093 } 5094 return 0; 5095 } 5096 5097 /* 5098 ** Part of sqlite3WhereEnd() will rewrite opcodes to reference the 5099 ** index rather than the main table. In SQLITE_DEBUG mode, we want 5100 ** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine 5101 ** does that. 5102 */ 5103 #ifndef SQLITE_DEBUG 5104 # define OpcodeRewriteTrace(D,K,P) /* no-op */ 5105 #else 5106 # define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P) 5107 static void sqlite3WhereOpcodeRewriteTrace( 5108 sqlite3 *db, 5109 int pc, 5110 VdbeOp *pOp 5111 ){ 5112 if( (db->flags & SQLITE_VdbeAddopTrace)==0 ) return; 5113 sqlite3VdbePrintOp(0, pc, pOp); 5114 } 5115 #endif 5116 5117 /* 5118 ** Generate the end of the WHERE loop. See comments on 5119 ** sqlite3WhereBegin() for additional information. 5120 */ 5121 void sqlite3WhereEnd(WhereInfo *pWInfo){ 5122 Parse *pParse = pWInfo->pParse; 5123 Vdbe *v = pParse->pVdbe; 5124 int i; 5125 WhereLevel *pLevel; 5126 WhereLoop *pLoop; 5127 SrcList *pTabList = pWInfo->pTabList; 5128 sqlite3 *db = pParse->db; 5129 5130 /* Generate loop termination code. 5131 */ 5132 VdbeModuleComment((v, "End WHERE-core")); 5133 for(i=pWInfo->nLevel-1; i>=0; i--){ 5134 int addr; 5135 pLevel = &pWInfo->a[i]; 5136 pLoop = pLevel->pWLoop; 5137 if( pLevel->op!=OP_Noop ){ 5138 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5139 int addrSeek = 0; 5140 Index *pIdx; 5141 int n; 5142 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 5143 && i==pWInfo->nLevel-1 /* Ticket [ef9318757b152e3] 2017-10-21 */ 5144 && (pLoop->wsFlags & WHERE_INDEXED)!=0 5145 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 5146 && (n = pLoop->u.btree.nIdxCol)>0 5147 && pIdx->aiRowLogEst[n]>=36 5148 ){ 5149 int r1 = pParse->nMem+1; 5150 int j, op; 5151 for(j=0; j<n; j++){ 5152 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 5153 } 5154 pParse->nMem += n+1; 5155 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 5156 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 5157 VdbeCoverageIf(v, op==OP_SeekLT); 5158 VdbeCoverageIf(v, op==OP_SeekGT); 5159 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 5160 } 5161 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 5162 /* The common case: Advance to the next row */ 5163 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5164 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 5165 sqlite3VdbeChangeP5(v, pLevel->p5); 5166 VdbeCoverage(v); 5167 VdbeCoverageIf(v, pLevel->op==OP_Next); 5168 VdbeCoverageIf(v, pLevel->op==OP_Prev); 5169 VdbeCoverageIf(v, pLevel->op==OP_VNext); 5170 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5171 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 5172 #endif 5173 }else{ 5174 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5175 } 5176 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 5177 struct InLoop *pIn; 5178 int j; 5179 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 5180 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 5181 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5182 if( pIn->eEndLoopOp!=OP_Noop ){ 5183 if( pIn->nPrefix ){ 5184 assert( pLoop->wsFlags & WHERE_IN_EARLYOUT ); 5185 sqlite3VdbeAddOp4Int(v, OP_IfNoHope, pLevel->iIdxCur, 5186 sqlite3VdbeCurrentAddr(v)+2, 5187 pIn->iBase, pIn->nPrefix); 5188 VdbeCoverage(v); 5189 } 5190 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 5191 VdbeCoverage(v); 5192 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Prev); 5193 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Next); 5194 } 5195 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 5196 } 5197 } 5198 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 5199 if( pLevel->addrSkip ){ 5200 sqlite3VdbeGoto(v, pLevel->addrSkip); 5201 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 5202 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 5203 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 5204 } 5205 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 5206 if( pLevel->addrLikeRep ){ 5207 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 5208 pLevel->addrLikeRep); 5209 VdbeCoverage(v); 5210 } 5211 #endif 5212 if( pLevel->iLeftJoin ){ 5213 int ws = pLoop->wsFlags; 5214 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 5215 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 5216 if( (ws & WHERE_IDX_ONLY)==0 ){ 5217 assert( pLevel->iTabCur==pTabList->a[pLevel->iFrom].iCursor ); 5218 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iTabCur); 5219 } 5220 if( (ws & WHERE_INDEXED) 5221 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCovidx) 5222 ){ 5223 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 5224 } 5225 if( pLevel->op==OP_Return ){ 5226 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 5227 }else{ 5228 sqlite3VdbeGoto(v, pLevel->addrFirst); 5229 } 5230 sqlite3VdbeJumpHere(v, addr); 5231 } 5232 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 5233 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 5234 } 5235 5236 /* The "break" point is here, just past the end of the outer loop. 5237 ** Set it. 5238 */ 5239 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5240 5241 assert( pWInfo->nLevel<=pTabList->nSrc ); 5242 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5243 int k, last; 5244 VdbeOp *pOp; 5245 Index *pIdx = 0; 5246 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; 5247 Table *pTab = pTabItem->pTab; 5248 assert( pTab!=0 ); 5249 pLoop = pLevel->pWLoop; 5250 5251 /* For a co-routine, change all OP_Column references to the table of 5252 ** the co-routine into OP_Copy of result contained in a register. 5253 ** OP_Rowid becomes OP_Null. 5254 */ 5255 if( pTabItem->fg.viaCoroutine ){ 5256 testcase( pParse->db->mallocFailed ); 5257 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 5258 pTabItem->regResult, 0); 5259 continue; 5260 } 5261 5262 /* If this scan uses an index, make VDBE code substitutions to read data 5263 ** from the index instead of from the table where possible. In some cases 5264 ** this optimization prevents the table from ever being read, which can 5265 ** yield a significant performance boost. 5266 ** 5267 ** Calls to the code generator in between sqlite3WhereBegin and 5268 ** sqlite3WhereEnd will have created code that references the table 5269 ** directly. This loop scans all that code looking for opcodes 5270 ** that reference the table and converts them into opcodes that 5271 ** reference the index. 5272 */ 5273 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 5274 pIdx = pLoop->u.btree.pIndex; 5275 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 5276 pIdx = pLevel->u.pCovidx; 5277 } 5278 if( pIdx 5279 && (pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable)) 5280 && !db->mallocFailed 5281 ){ 5282 last = sqlite3VdbeCurrentAddr(v); 5283 k = pLevel->addrBody; 5284 #ifdef SQLITE_DEBUG 5285 if( db->flags & SQLITE_VdbeAddopTrace ){ 5286 printf("TRANSLATE opcodes in range %d..%d\n", k, last-1); 5287 } 5288 #endif 5289 pOp = sqlite3VdbeGetOp(v, k); 5290 for(; k<last; k++, pOp++){ 5291 if( pOp->p1!=pLevel->iTabCur ) continue; 5292 if( pOp->opcode==OP_Column 5293 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC 5294 || pOp->opcode==OP_Offset 5295 #endif 5296 ){ 5297 int x = pOp->p2; 5298 assert( pIdx->pTable==pTab ); 5299 if( !HasRowid(pTab) ){ 5300 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 5301 x = pPk->aiColumn[x]; 5302 assert( x>=0 ); 5303 } 5304 x = sqlite3ColumnOfIndex(pIdx, x); 5305 if( x>=0 ){ 5306 pOp->p2 = x; 5307 pOp->p1 = pLevel->iIdxCur; 5308 OpcodeRewriteTrace(db, k, pOp); 5309 } 5310 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 5311 || pWInfo->eOnePass ); 5312 }else if( pOp->opcode==OP_Rowid ){ 5313 pOp->p1 = pLevel->iIdxCur; 5314 pOp->opcode = OP_IdxRowid; 5315 OpcodeRewriteTrace(db, k, pOp); 5316 }else if( pOp->opcode==OP_IfNullRow ){ 5317 pOp->p1 = pLevel->iIdxCur; 5318 OpcodeRewriteTrace(db, k, pOp); 5319 } 5320 } 5321 #ifdef SQLITE_DEBUG 5322 if( db->flags & SQLITE_VdbeAddopTrace ) printf("TRANSLATE complete\n"); 5323 #endif 5324 } 5325 } 5326 5327 /* Final cleanup 5328 */ 5329 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5330 whereInfoFree(db, pWInfo); 5331 return; 5332 } 5333