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 /* 41 ** Return the estimated number of output rows from a WHERE clause 42 */ 43 LogEst sqlite3WhereOutputRowCount(WhereInfo *pWInfo){ 44 return pWInfo->nRowOut; 45 } 46 47 /* 48 ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this 49 ** WHERE clause returns outputs for DISTINCT processing. 50 */ 51 int sqlite3WhereIsDistinct(WhereInfo *pWInfo){ 52 return pWInfo->eDistinct; 53 } 54 55 /* 56 ** Return the number of ORDER BY terms that are satisfied by the 57 ** WHERE clause. A return of 0 means that the output must be 58 ** completely sorted. A return equal to the number of ORDER BY 59 ** terms means that no sorting is needed at all. A return that 60 ** is positive but less than the number of ORDER BY terms means that 61 ** block sorting is required. 62 */ 63 int sqlite3WhereIsOrdered(WhereInfo *pWInfo){ 64 return pWInfo->nOBSat; 65 } 66 67 /* 68 ** In the ORDER BY LIMIT optimization, if the inner-most loop is known 69 ** to emit rows in increasing order, and if the last row emitted by the 70 ** inner-most loop did not fit within the sorter, then we can skip all 71 ** subsequent rows for the current iteration of the inner loop (because they 72 ** will not fit in the sorter either) and continue with the second inner 73 ** loop - the loop immediately outside the inner-most. 74 ** 75 ** When a row does not fit in the sorter (because the sorter already 76 ** holds LIMIT+OFFSET rows that are smaller), then a jump is made to the 77 ** label returned by this function. 78 ** 79 ** If the ORDER BY LIMIT optimization applies, the jump destination should 80 ** be the continuation for the second-inner-most loop. If the ORDER BY 81 ** LIMIT optimization does not apply, then the jump destination should 82 ** be the continuation for the inner-most loop. 83 ** 84 ** It is always safe for this routine to return the continuation of the 85 ** inner-most loop, in the sense that a correct answer will result. 86 ** Returning the continuation the second inner loop is an optimization 87 ** that might make the code run a little faster, but should not change 88 ** the final answer. 89 */ 90 int sqlite3WhereOrderByLimitOptLabel(WhereInfo *pWInfo){ 91 WhereLevel *pInner; 92 if( !pWInfo->bOrderedInnerLoop ){ 93 /* The ORDER BY LIMIT optimization does not apply. Jump to the 94 ** continuation of the inner-most loop. */ 95 return pWInfo->iContinue; 96 } 97 pInner = &pWInfo->a[pWInfo->nLevel-1]; 98 assert( pInner->addrNxt!=0 ); 99 return pInner->addrNxt; 100 } 101 102 /* 103 ** While generating code for the min/max optimization, after handling 104 ** the aggregate-step call to min() or max(), check to see if any 105 ** additional looping is required. If the output order is such that 106 ** we are certain that the correct answer has already been found, then 107 ** code an OP_Goto to by pass subsequent processing. 108 ** 109 ** Any extra OP_Goto that is coded here is an optimization. The 110 ** correct answer should be obtained regardless. This OP_Goto just 111 ** makes the answer appear faster. 112 */ 113 void sqlite3WhereMinMaxOptEarlyOut(Vdbe *v, WhereInfo *pWInfo){ 114 WhereLevel *pInner; 115 int i; 116 if( !pWInfo->bOrderedInnerLoop ) return; 117 if( pWInfo->nOBSat==0 ) return; 118 for(i=pWInfo->nLevel-1; i>=0; i--){ 119 pInner = &pWInfo->a[i]; 120 if( (pInner->pWLoop->wsFlags & WHERE_COLUMN_IN)!=0 ){ 121 sqlite3VdbeGoto(v, pInner->addrNxt); 122 return; 123 } 124 } 125 sqlite3VdbeGoto(v, pWInfo->iBreak); 126 } 127 128 /* 129 ** Return the VDBE address or label to jump to in order to continue 130 ** immediately with the next row of a WHERE clause. 131 */ 132 int sqlite3WhereContinueLabel(WhereInfo *pWInfo){ 133 assert( pWInfo->iContinue!=0 ); 134 return pWInfo->iContinue; 135 } 136 137 /* 138 ** Return the VDBE address or label to jump to in order to break 139 ** out of a WHERE loop. 140 */ 141 int sqlite3WhereBreakLabel(WhereInfo *pWInfo){ 142 return pWInfo->iBreak; 143 } 144 145 /* 146 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to 147 ** operate directly on the rowids returned by a WHERE clause. Return 148 ** ONEPASS_SINGLE (1) if the statement can operation directly because only 149 ** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass 150 ** optimization can be used on multiple 151 ** 152 ** If the ONEPASS optimization is used (if this routine returns true) 153 ** then also write the indices of open cursors used by ONEPASS 154 ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data 155 ** table and iaCur[1] gets the cursor used by an auxiliary index. 156 ** Either value may be -1, indicating that cursor is not used. 157 ** Any cursors returned will have been opened for writing. 158 ** 159 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is 160 ** unable to use the ONEPASS optimization. 161 */ 162 int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){ 163 memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2); 164 #ifdef WHERETRACE_ENABLED 165 if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){ 166 sqlite3DebugPrintf("%s cursors: %d %d\n", 167 pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI", 168 aiCur[0], aiCur[1]); 169 } 170 #endif 171 return pWInfo->eOnePass; 172 } 173 174 /* 175 ** Return TRUE if the WHERE loop uses the OP_DeferredSeek opcode to move 176 ** the data cursor to the row selected by the index cursor. 177 */ 178 int sqlite3WhereUsesDeferredSeek(WhereInfo *pWInfo){ 179 return pWInfo->bDeferredSeek; 180 } 181 182 /* 183 ** Move the content of pSrc into pDest 184 */ 185 static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){ 186 pDest->n = pSrc->n; 187 memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0])); 188 } 189 190 /* 191 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet. 192 ** 193 ** The new entry might overwrite an existing entry, or it might be 194 ** appended, or it might be discarded. Do whatever is the right thing 195 ** so that pSet keeps the N_OR_COST best entries seen so far. 196 */ 197 static int whereOrInsert( 198 WhereOrSet *pSet, /* The WhereOrSet to be updated */ 199 Bitmask prereq, /* Prerequisites of the new entry */ 200 LogEst rRun, /* Run-cost of the new entry */ 201 LogEst nOut /* Number of outputs for the new entry */ 202 ){ 203 u16 i; 204 WhereOrCost *p; 205 for(i=pSet->n, p=pSet->a; i>0; i--, p++){ 206 if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){ 207 goto whereOrInsert_done; 208 } 209 if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){ 210 return 0; 211 } 212 } 213 if( pSet->n<N_OR_COST ){ 214 p = &pSet->a[pSet->n++]; 215 p->nOut = nOut; 216 }else{ 217 p = pSet->a; 218 for(i=1; i<pSet->n; i++){ 219 if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i; 220 } 221 if( p->rRun<=rRun ) return 0; 222 } 223 whereOrInsert_done: 224 p->prereq = prereq; 225 p->rRun = rRun; 226 if( p->nOut>nOut ) p->nOut = nOut; 227 return 1; 228 } 229 230 /* 231 ** Return the bitmask for the given cursor number. Return 0 if 232 ** iCursor is not in the set. 233 */ 234 Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){ 235 int i; 236 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 ); 237 for(i=0; i<pMaskSet->n; i++){ 238 if( pMaskSet->ix[i]==iCursor ){ 239 return MASKBIT(i); 240 } 241 } 242 return 0; 243 } 244 245 /* 246 ** Create a new mask for cursor iCursor. 247 ** 248 ** There is one cursor per table in the FROM clause. The number of 249 ** tables in the FROM clause is limited by a test early in the 250 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] 251 ** array will never overflow. 252 */ 253 static void createMask(WhereMaskSet *pMaskSet, int iCursor){ 254 assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); 255 pMaskSet->ix[pMaskSet->n++] = iCursor; 256 } 257 258 /* 259 ** If the right-hand branch of the expression is a TK_COLUMN, then return 260 ** a pointer to the right-hand branch. Otherwise, return NULL. 261 */ 262 static Expr *whereRightSubexprIsColumn(Expr *p){ 263 p = sqlite3ExprSkipCollateAndLikely(p->pRight); 264 if( ALWAYS(p!=0) && p->op==TK_COLUMN ) return p; 265 return 0; 266 } 267 268 /* 269 ** Advance to the next WhereTerm that matches according to the criteria 270 ** established when the pScan object was initialized by whereScanInit(). 271 ** Return NULL if there are no more matching WhereTerms. 272 */ 273 static WhereTerm *whereScanNext(WhereScan *pScan){ 274 int iCur; /* The cursor on the LHS of the term */ 275 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */ 276 Expr *pX; /* An expression being tested */ 277 WhereClause *pWC; /* Shorthand for pScan->pWC */ 278 WhereTerm *pTerm; /* The term being tested */ 279 int k = pScan->k; /* Where to start scanning */ 280 281 assert( pScan->iEquiv<=pScan->nEquiv ); 282 pWC = pScan->pWC; 283 while(1){ 284 iColumn = pScan->aiColumn[pScan->iEquiv-1]; 285 iCur = pScan->aiCur[pScan->iEquiv-1]; 286 assert( pWC!=0 ); 287 do{ 288 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){ 289 if( pTerm->leftCursor==iCur 290 && pTerm->u.x.leftColumn==iColumn 291 && (iColumn!=XN_EXPR 292 || sqlite3ExprCompareSkip(pTerm->pExpr->pLeft, 293 pScan->pIdxExpr,iCur)==0) 294 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin)) 295 ){ 296 if( (pTerm->eOperator & WO_EQUIV)!=0 297 && pScan->nEquiv<ArraySize(pScan->aiCur) 298 && (pX = whereRightSubexprIsColumn(pTerm->pExpr))!=0 299 ){ 300 int j; 301 for(j=0; j<pScan->nEquiv; j++){ 302 if( pScan->aiCur[j]==pX->iTable 303 && pScan->aiColumn[j]==pX->iColumn ){ 304 break; 305 } 306 } 307 if( j==pScan->nEquiv ){ 308 pScan->aiCur[j] = pX->iTable; 309 pScan->aiColumn[j] = pX->iColumn; 310 pScan->nEquiv++; 311 } 312 } 313 if( (pTerm->eOperator & pScan->opMask)!=0 ){ 314 /* Verify the affinity and collating sequence match */ 315 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){ 316 CollSeq *pColl; 317 Parse *pParse = pWC->pWInfo->pParse; 318 pX = pTerm->pExpr; 319 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){ 320 continue; 321 } 322 assert(pX->pLeft); 323 pColl = sqlite3ExprCompareCollSeq(pParse, pX); 324 if( pColl==0 ) pColl = pParse->db->pDfltColl; 325 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){ 326 continue; 327 } 328 } 329 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0 330 && (pX = pTerm->pExpr->pRight)->op==TK_COLUMN 331 && pX->iTable==pScan->aiCur[0] 332 && pX->iColumn==pScan->aiColumn[0] 333 ){ 334 testcase( pTerm->eOperator & WO_IS ); 335 continue; 336 } 337 pScan->pWC = pWC; 338 pScan->k = k+1; 339 return pTerm; 340 } 341 } 342 } 343 pWC = pWC->pOuter; 344 k = 0; 345 }while( pWC!=0 ); 346 if( pScan->iEquiv>=pScan->nEquiv ) break; 347 pWC = pScan->pOrigWC; 348 k = 0; 349 pScan->iEquiv++; 350 } 351 return 0; 352 } 353 354 /* 355 ** This is whereScanInit() for the case of an index on an expression. 356 ** It is factored out into a separate tail-recursion subroutine so that 357 ** the normal whereScanInit() routine, which is a high-runner, does not 358 ** need to push registers onto the stack as part of its prologue. 359 */ 360 static SQLITE_NOINLINE WhereTerm *whereScanInitIndexExpr(WhereScan *pScan){ 361 pScan->idxaff = sqlite3ExprAffinity(pScan->pIdxExpr); 362 return whereScanNext(pScan); 363 } 364 365 /* 366 ** Initialize a WHERE clause scanner object. Return a pointer to the 367 ** first match. Return NULL if there are no matches. 368 ** 369 ** The scanner will be searching the WHERE clause pWC. It will look 370 ** for terms of the form "X <op> <expr>" where X is column iColumn of table 371 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx 372 ** must be one of the indexes of table iCur. 373 ** 374 ** The <op> must be one of the operators described by opMask. 375 ** 376 ** If the search is for X and the WHERE clause contains terms of the 377 ** form X=Y then this routine might also return terms of the form 378 ** "Y <op> <expr>". The number of levels of transitivity is limited, 379 ** but is enough to handle most commonly occurring SQL statements. 380 ** 381 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with 382 ** index pIdx. 383 */ 384 static WhereTerm *whereScanInit( 385 WhereScan *pScan, /* The WhereScan object being initialized */ 386 WhereClause *pWC, /* The WHERE clause to be scanned */ 387 int iCur, /* Cursor to scan for */ 388 int iColumn, /* Column to scan for */ 389 u32 opMask, /* Operator(s) to scan for */ 390 Index *pIdx /* Must be compatible with this index */ 391 ){ 392 pScan->pOrigWC = pWC; 393 pScan->pWC = pWC; 394 pScan->pIdxExpr = 0; 395 pScan->idxaff = 0; 396 pScan->zCollName = 0; 397 pScan->opMask = opMask; 398 pScan->k = 0; 399 pScan->aiCur[0] = iCur; 400 pScan->nEquiv = 1; 401 pScan->iEquiv = 1; 402 if( pIdx ){ 403 int j = iColumn; 404 iColumn = pIdx->aiColumn[j]; 405 if( iColumn==XN_EXPR ){ 406 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr; 407 pScan->zCollName = pIdx->azColl[j]; 408 pScan->aiColumn[0] = XN_EXPR; 409 return whereScanInitIndexExpr(pScan); 410 }else if( iColumn==pIdx->pTable->iPKey ){ 411 iColumn = XN_ROWID; 412 }else if( iColumn>=0 ){ 413 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; 414 pScan->zCollName = pIdx->azColl[j]; 415 } 416 }else if( iColumn==XN_EXPR ){ 417 return 0; 418 } 419 pScan->aiColumn[0] = iColumn; 420 return whereScanNext(pScan); 421 } 422 423 /* 424 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 425 ** where X is a reference to the iColumn of table iCur or of index pIdx 426 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by 427 ** the op parameter. Return a pointer to the term. Return 0 if not found. 428 ** 429 ** If pIdx!=0 then it must be one of the indexes of table iCur. 430 ** Search for terms matching the iColumn-th column of pIdx 431 ** rather than the iColumn-th column of table iCur. 432 ** 433 ** The term returned might by Y=<expr> if there is another constraint in 434 ** the WHERE clause that specifies that X=Y. Any such constraints will be 435 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 436 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11 437 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10 438 ** other equivalent values. Hence a search for X will return <expr> if X=A1 439 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>. 440 ** 441 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 442 ** then try for the one with no dependencies on <expr> - in other words where 443 ** <expr> is a constant expression of some kind. Only return entries of 444 ** the form "X <op> Y" where Y is a column in another table if no terms of 445 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 446 ** exist, try to return a term that does not use WO_EQUIV. 447 */ 448 WhereTerm *sqlite3WhereFindTerm( 449 WhereClause *pWC, /* The WHERE clause to be searched */ 450 int iCur, /* Cursor number of LHS */ 451 int iColumn, /* Column number of LHS */ 452 Bitmask notReady, /* RHS must not overlap with this mask */ 453 u32 op, /* Mask of WO_xx values describing operator */ 454 Index *pIdx /* Must be compatible with this index, if not NULL */ 455 ){ 456 WhereTerm *pResult = 0; 457 WhereTerm *p; 458 WhereScan scan; 459 460 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); 461 op &= WO_EQ|WO_IS; 462 while( p ){ 463 if( (p->prereqRight & notReady)==0 ){ 464 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){ 465 testcase( p->eOperator & WO_IS ); 466 return p; 467 } 468 if( pResult==0 ) pResult = p; 469 } 470 p = whereScanNext(&scan); 471 } 472 return pResult; 473 } 474 475 /* 476 ** This function searches pList for an entry that matches the iCol-th column 477 ** of index pIdx. 478 ** 479 ** If such an expression is found, its index in pList->a[] is returned. If 480 ** no expression is found, -1 is returned. 481 */ 482 static int findIndexCol( 483 Parse *pParse, /* Parse context */ 484 ExprList *pList, /* Expression list to search */ 485 int iBase, /* Cursor for table associated with pIdx */ 486 Index *pIdx, /* Index to match column of */ 487 int iCol /* Column of index to match */ 488 ){ 489 int i; 490 const char *zColl = pIdx->azColl[iCol]; 491 492 for(i=0; i<pList->nExpr; i++){ 493 Expr *p = sqlite3ExprSkipCollateAndLikely(pList->a[i].pExpr); 494 if( ALWAYS(p!=0) 495 && (p->op==TK_COLUMN || p->op==TK_AGG_COLUMN) 496 && p->iColumn==pIdx->aiColumn[iCol] 497 && p->iTable==iBase 498 ){ 499 CollSeq *pColl = sqlite3ExprNNCollSeq(pParse, pList->a[i].pExpr); 500 if( 0==sqlite3StrICmp(pColl->zName, zColl) ){ 501 return i; 502 } 503 } 504 } 505 506 return -1; 507 } 508 509 /* 510 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL 511 */ 512 static int indexColumnNotNull(Index *pIdx, int iCol){ 513 int j; 514 assert( pIdx!=0 ); 515 assert( iCol>=0 && iCol<pIdx->nColumn ); 516 j = pIdx->aiColumn[iCol]; 517 if( j>=0 ){ 518 return pIdx->pTable->aCol[j].notNull; 519 }else if( j==(-1) ){ 520 return 1; 521 }else{ 522 assert( j==(-2) ); 523 return 0; /* Assume an indexed expression can always yield a NULL */ 524 525 } 526 } 527 528 /* 529 ** Return true if the DISTINCT expression-list passed as the third argument 530 ** is redundant. 531 ** 532 ** A DISTINCT list is redundant if any subset of the columns in the 533 ** DISTINCT list are collectively unique and individually non-null. 534 */ 535 static int isDistinctRedundant( 536 Parse *pParse, /* Parsing context */ 537 SrcList *pTabList, /* The FROM clause */ 538 WhereClause *pWC, /* The WHERE clause */ 539 ExprList *pDistinct /* The result set that needs to be DISTINCT */ 540 ){ 541 Table *pTab; 542 Index *pIdx; 543 int i; 544 int iBase; 545 546 /* If there is more than one table or sub-select in the FROM clause of 547 ** this query, then it will not be possible to show that the DISTINCT 548 ** clause is redundant. */ 549 if( pTabList->nSrc!=1 ) return 0; 550 iBase = pTabList->a[0].iCursor; 551 pTab = pTabList->a[0].pTab; 552 553 /* If any of the expressions is an IPK column on table iBase, then return 554 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 555 ** current SELECT is a correlated sub-query. 556 */ 557 for(i=0; i<pDistinct->nExpr; i++){ 558 Expr *p = sqlite3ExprSkipCollateAndLikely(pDistinct->a[i].pExpr); 559 if( NEVER(p==0) ) continue; 560 if( p->op!=TK_COLUMN && p->op!=TK_AGG_COLUMN ) continue; 561 if( p->iTable==iBase && p->iColumn<0 ) return 1; 562 } 563 564 /* Loop through all indices on the table, checking each to see if it makes 565 ** the DISTINCT qualifier redundant. It does so if: 566 ** 567 ** 1. The index is itself UNIQUE, and 568 ** 569 ** 2. All of the columns in the index are either part of the pDistinct 570 ** list, or else the WHERE clause contains a term of the form "col=X", 571 ** where X is a constant value. The collation sequences of the 572 ** comparison and select-list expressions must match those of the index. 573 ** 574 ** 3. All of those index columns for which the WHERE clause does not 575 ** contain a "col=X" term are subject to a NOT NULL constraint. 576 */ 577 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 578 if( !IsUniqueIndex(pIdx) ) continue; 579 if( pIdx->pPartIdxWhere ) continue; 580 for(i=0; i<pIdx->nKeyCol; i++){ 581 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){ 582 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break; 583 if( indexColumnNotNull(pIdx, i)==0 ) break; 584 } 585 } 586 if( i==pIdx->nKeyCol ){ 587 /* This index implies that the DISTINCT qualifier is redundant. */ 588 return 1; 589 } 590 } 591 592 return 0; 593 } 594 595 596 /* 597 ** Estimate the logarithm of the input value to base 2. 598 */ 599 static LogEst estLog(LogEst N){ 600 return N<=10 ? 0 : sqlite3LogEst(N) - 33; 601 } 602 603 /* 604 ** Convert OP_Column opcodes to OP_Copy in previously generated code. 605 ** 606 ** This routine runs over generated VDBE code and translates OP_Column 607 ** opcodes into OP_Copy when the table is being accessed via co-routine 608 ** instead of via table lookup. 609 ** 610 ** If the iAutoidxCur is not zero, then any OP_Rowid instructions on 611 ** cursor iTabCur are transformed into OP_Sequence opcode for the 612 ** iAutoidxCur cursor, in order to generate unique rowids for the 613 ** automatic index being generated. 614 */ 615 static void translateColumnToCopy( 616 Parse *pParse, /* Parsing context */ 617 int iStart, /* Translate from this opcode to the end */ 618 int iTabCur, /* OP_Column/OP_Rowid references to this table */ 619 int iRegister, /* The first column is in this register */ 620 int iAutoidxCur /* If non-zero, cursor of autoindex being generated */ 621 ){ 622 Vdbe *v = pParse->pVdbe; 623 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart); 624 int iEnd = sqlite3VdbeCurrentAddr(v); 625 if( pParse->db->mallocFailed ) return; 626 for(; iStart<iEnd; iStart++, pOp++){ 627 if( pOp->p1!=iTabCur ) continue; 628 if( pOp->opcode==OP_Column ){ 629 pOp->opcode = OP_Copy; 630 pOp->p1 = pOp->p2 + iRegister; 631 pOp->p2 = pOp->p3; 632 pOp->p3 = 0; 633 }else if( pOp->opcode==OP_Rowid ){ 634 pOp->opcode = OP_Sequence; 635 pOp->p1 = iAutoidxCur; 636 #ifdef SQLITE_ALLOW_ROWID_IN_VIEW 637 if( iAutoidxCur==0 ){ 638 pOp->opcode = OP_Null; 639 pOp->p3 = 0; 640 } 641 #endif 642 } 643 } 644 } 645 646 /* 647 ** Two routines for printing the content of an sqlite3_index_info 648 ** structure. Used for testing and debugging only. If neither 649 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 650 ** are no-ops. 651 */ 652 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) 653 static void whereTraceIndexInfoInputs(sqlite3_index_info *p){ 654 int i; 655 if( !sqlite3WhereTrace ) return; 656 for(i=0; i<p->nConstraint; i++){ 657 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", 658 i, 659 p->aConstraint[i].iColumn, 660 p->aConstraint[i].iTermOffset, 661 p->aConstraint[i].op, 662 p->aConstraint[i].usable); 663 } 664 for(i=0; i<p->nOrderBy; i++){ 665 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 666 i, 667 p->aOrderBy[i].iColumn, 668 p->aOrderBy[i].desc); 669 } 670 } 671 static void whereTraceIndexInfoOutputs(sqlite3_index_info *p){ 672 int i; 673 if( !sqlite3WhereTrace ) return; 674 for(i=0; i<p->nConstraint; i++){ 675 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 676 i, 677 p->aConstraintUsage[i].argvIndex, 678 p->aConstraintUsage[i].omit); 679 } 680 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 681 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 682 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 683 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 684 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); 685 } 686 #else 687 #define whereTraceIndexInfoInputs(A) 688 #define whereTraceIndexInfoOutputs(A) 689 #endif 690 691 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 692 /* 693 ** Return TRUE if the WHERE clause term pTerm is of a form where it 694 ** could be used with an index to access pSrc, assuming an appropriate 695 ** index existed. 696 */ 697 static int termCanDriveIndex( 698 WhereTerm *pTerm, /* WHERE clause term to check */ 699 SrcItem *pSrc, /* Table we are trying to access */ 700 Bitmask notReady /* Tables in outer loops of the join */ 701 ){ 702 char aff; 703 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 704 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0; 705 if( (pSrc->fg.jointype & JT_LEFT) 706 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 707 && (pTerm->eOperator & WO_IS) 708 ){ 709 /* Cannot use an IS term from the WHERE clause as an index driver for 710 ** the RHS of a LEFT JOIN. Such a term can only be used if it is from 711 ** the ON clause. */ 712 return 0; 713 } 714 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 715 if( pTerm->u.x.leftColumn<0 ) return 0; 716 aff = pSrc->pTab->aCol[pTerm->u.x.leftColumn].affinity; 717 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 718 testcase( pTerm->pExpr->op==TK_IS ); 719 return 1; 720 } 721 #endif 722 723 724 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 725 /* 726 ** Generate code to construct the Index object for an automatic index 727 ** and to set up the WhereLevel object pLevel so that the code generator 728 ** makes use of the automatic index. 729 */ 730 static void constructAutomaticIndex( 731 Parse *pParse, /* The parsing context */ 732 WhereClause *pWC, /* The WHERE clause */ 733 SrcItem *pSrc, /* The FROM clause term to get the next index */ 734 Bitmask notReady, /* Mask of cursors that are not available */ 735 WhereLevel *pLevel /* Write new index here */ 736 ){ 737 int nKeyCol; /* Number of columns in the constructed index */ 738 WhereTerm *pTerm; /* A single term of the WHERE clause */ 739 WhereTerm *pWCEnd; /* End of pWC->a[] */ 740 Index *pIdx; /* Object describing the transient index */ 741 Vdbe *v; /* Prepared statement under construction */ 742 int addrInit; /* Address of the initialization bypass jump */ 743 Table *pTable; /* The table being indexed */ 744 int addrTop; /* Top of the index fill loop */ 745 int regRecord; /* Register holding an index record */ 746 int n; /* Column counter */ 747 int i; /* Loop counter */ 748 int mxBitCol; /* Maximum column in pSrc->colUsed */ 749 CollSeq *pColl; /* Collating sequence to on a column */ 750 WhereLoop *pLoop; /* The Loop object */ 751 char *zNotUsed; /* Extra space on the end of pIdx */ 752 Bitmask idxCols; /* Bitmap of columns used for indexing */ 753 Bitmask extraCols; /* Bitmap of additional columns */ 754 u8 sentWarning = 0; /* True if a warnning has been issued */ 755 Expr *pPartial = 0; /* Partial Index Expression */ 756 int iContinue = 0; /* Jump here to skip excluded rows */ 757 SrcItem *pTabItem; /* FROM clause term being indexed */ 758 int addrCounter = 0; /* Address where integer counter is initialized */ 759 int regBase; /* Array of registers where record is assembled */ 760 761 /* Generate code to skip over the creation and initialization of the 762 ** transient index on 2nd and subsequent iterations of the loop. */ 763 v = pParse->pVdbe; 764 assert( v!=0 ); 765 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 766 767 /* Count the number of columns that will be added to the index 768 ** and used to match WHERE clause constraints */ 769 nKeyCol = 0; 770 pTable = pSrc->pTab; 771 pWCEnd = &pWC->a[pWC->nTerm]; 772 pLoop = pLevel->pWLoop; 773 idxCols = 0; 774 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 775 Expr *pExpr = pTerm->pExpr; 776 assert( !ExprHasProperty(pExpr, EP_FromJoin) /* prereq always non-zero */ 777 || pExpr->iRightJoinTable!=pSrc->iCursor /* for the right-hand */ 778 || pLoop->prereq!=0 ); /* table of a LEFT JOIN */ 779 if( pLoop->prereq==0 780 && (pTerm->wtFlags & TERM_VIRTUAL)==0 781 && !ExprHasProperty(pExpr, EP_FromJoin) 782 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) ){ 783 pPartial = sqlite3ExprAnd(pParse, pPartial, 784 sqlite3ExprDup(pParse->db, pExpr, 0)); 785 } 786 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 787 int iCol = pTerm->u.x.leftColumn; 788 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 789 testcase( iCol==BMS ); 790 testcase( iCol==BMS-1 ); 791 if( !sentWarning ){ 792 sqlite3_log(SQLITE_WARNING_AUTOINDEX, 793 "automatic index on %s(%s)", pTable->zName, 794 pTable->aCol[iCol].zName); 795 sentWarning = 1; 796 } 797 if( (idxCols & cMask)==0 ){ 798 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ 799 goto end_auto_index_create; 800 } 801 pLoop->aLTerm[nKeyCol++] = pTerm; 802 idxCols |= cMask; 803 } 804 } 805 } 806 assert( nKeyCol>0 || pParse->db->mallocFailed ); 807 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; 808 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED 809 | WHERE_AUTO_INDEX; 810 811 /* Count the number of additional columns needed to create a 812 ** covering index. A "covering index" is an index that contains all 813 ** columns that are needed by the query. With a covering index, the 814 ** original table never needs to be accessed. Automatic indices must 815 ** be a covering index because the index will not be updated if the 816 ** original table changes and the index and table cannot both be used 817 ** if they go out of sync. 818 */ 819 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); 820 mxBitCol = MIN(BMS-1,pTable->nCol); 821 testcase( pTable->nCol==BMS-1 ); 822 testcase( pTable->nCol==BMS-2 ); 823 for(i=0; i<mxBitCol; i++){ 824 if( extraCols & MASKBIT(i) ) nKeyCol++; 825 } 826 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 827 nKeyCol += pTable->nCol - BMS + 1; 828 } 829 830 /* Construct the Index object to describe this index */ 831 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); 832 if( pIdx==0 ) goto end_auto_index_create; 833 pLoop->u.btree.pIndex = pIdx; 834 pIdx->zName = "auto-index"; 835 pIdx->pTable = pTable; 836 n = 0; 837 idxCols = 0; 838 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 839 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 840 int iCol = pTerm->u.x.leftColumn; 841 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 842 testcase( iCol==BMS-1 ); 843 testcase( iCol==BMS ); 844 if( (idxCols & cMask)==0 ){ 845 Expr *pX = pTerm->pExpr; 846 idxCols |= cMask; 847 pIdx->aiColumn[n] = pTerm->u.x.leftColumn; 848 pColl = sqlite3ExprCompareCollSeq(pParse, pX); 849 assert( pColl!=0 || pParse->nErr>0 ); /* TH3 collate01.800 */ 850 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY; 851 n++; 852 } 853 } 854 } 855 assert( (u32)n==pLoop->u.btree.nEq ); 856 857 /* Add additional columns needed to make the automatic index into 858 ** a covering index */ 859 for(i=0; i<mxBitCol; i++){ 860 if( extraCols & MASKBIT(i) ){ 861 pIdx->aiColumn[n] = i; 862 pIdx->azColl[n] = sqlite3StrBINARY; 863 n++; 864 } 865 } 866 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 867 for(i=BMS-1; i<pTable->nCol; i++){ 868 pIdx->aiColumn[n] = i; 869 pIdx->azColl[n] = sqlite3StrBINARY; 870 n++; 871 } 872 } 873 assert( n==nKeyCol ); 874 pIdx->aiColumn[n] = XN_ROWID; 875 pIdx->azColl[n] = sqlite3StrBINARY; 876 877 /* Create the automatic index */ 878 assert( pLevel->iIdxCur>=0 ); 879 pLevel->iIdxCur = pParse->nTab++; 880 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); 881 sqlite3VdbeSetP4KeyInfo(pParse, pIdx); 882 VdbeComment((v, "for %s", pTable->zName)); 883 884 /* Fill the automatic index with content */ 885 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom]; 886 if( pTabItem->fg.viaCoroutine ){ 887 int regYield = pTabItem->regReturn; 888 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0); 889 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); 890 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 891 VdbeCoverage(v); 892 VdbeComment((v, "next row of %s", pTabItem->pTab->zName)); 893 }else{ 894 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); 895 } 896 if( pPartial ){ 897 iContinue = sqlite3VdbeMakeLabel(pParse); 898 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); 899 pLoop->wsFlags |= WHERE_PARTIALIDX; 900 } 901 regRecord = sqlite3GetTempReg(pParse); 902 regBase = sqlite3GenerateIndexKey( 903 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0 904 ); 905 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 906 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 907 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); 908 if( pTabItem->fg.viaCoroutine ){ 909 sqlite3VdbeChangeP2(v, addrCounter, regBase+n); 910 testcase( pParse->db->mallocFailed ); 911 assert( pLevel->iIdxCur>0 ); 912 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur, 913 pTabItem->regResult, pLevel->iIdxCur); 914 sqlite3VdbeGoto(v, addrTop); 915 pTabItem->fg.viaCoroutine = 0; 916 }else{ 917 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); 918 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 919 } 920 sqlite3VdbeJumpHere(v, addrTop); 921 sqlite3ReleaseTempReg(pParse, regRecord); 922 923 /* Jump here when skipping the initialization */ 924 sqlite3VdbeJumpHere(v, addrInit); 925 926 end_auto_index_create: 927 sqlite3ExprDelete(pParse->db, pPartial); 928 } 929 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 930 931 #ifndef SQLITE_OMIT_VIRTUALTABLE 932 /* 933 ** Allocate and populate an sqlite3_index_info structure. It is the 934 ** responsibility of the caller to eventually release the structure 935 ** by passing the pointer returned by this function to sqlite3_free(). 936 */ 937 static sqlite3_index_info *allocateIndexInfo( 938 Parse *pParse, /* The parsing context */ 939 WhereClause *pWC, /* The WHERE clause being analyzed */ 940 Bitmask mUnusable, /* Ignore terms with these prereqs */ 941 SrcItem *pSrc, /* The FROM clause term that is the vtab */ 942 ExprList *pOrderBy, /* The ORDER BY clause */ 943 u16 *pmNoOmit /* Mask of terms not to omit */ 944 ){ 945 int i, j; 946 int nTerm; 947 struct sqlite3_index_constraint *pIdxCons; 948 struct sqlite3_index_orderby *pIdxOrderBy; 949 struct sqlite3_index_constraint_usage *pUsage; 950 struct HiddenIndexInfo *pHidden; 951 WhereTerm *pTerm; 952 int nOrderBy; 953 sqlite3_index_info *pIdxInfo; 954 u16 mNoOmit = 0; 955 956 /* Count the number of possible WHERE clause constraints referring 957 ** to this virtual table */ 958 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 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_ISNULL ); 964 testcase( pTerm->eOperator & WO_IS ); 965 testcase( pTerm->eOperator & WO_ALL ); 966 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 967 if( pTerm->wtFlags & TERM_VNULL ) continue; 968 assert( pTerm->u.x.leftColumn>=(-1) ); 969 nTerm++; 970 } 971 972 /* If the ORDER BY clause contains only columns in the current 973 ** virtual table then allocate space for the aOrderBy part of 974 ** the sqlite3_index_info structure. 975 */ 976 nOrderBy = 0; 977 if( pOrderBy ){ 978 int n = pOrderBy->nExpr; 979 for(i=0; i<n; i++){ 980 Expr *pExpr = pOrderBy->a[i].pExpr; 981 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; 982 if( pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL ) break; 983 } 984 if( i==n){ 985 nOrderBy = n; 986 } 987 } 988 989 /* Allocate the sqlite3_index_info structure 990 */ 991 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 992 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 993 + sizeof(*pIdxOrderBy)*nOrderBy + sizeof(*pHidden) ); 994 if( pIdxInfo==0 ){ 995 sqlite3ErrorMsg(pParse, "out of memory"); 996 return 0; 997 } 998 pHidden = (struct HiddenIndexInfo*)&pIdxInfo[1]; 999 pIdxCons = (struct sqlite3_index_constraint*)&pHidden[1]; 1000 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 1001 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 1002 pIdxInfo->nOrderBy = nOrderBy; 1003 pIdxInfo->aConstraint = pIdxCons; 1004 pIdxInfo->aOrderBy = pIdxOrderBy; 1005 pIdxInfo->aConstraintUsage = pUsage; 1006 pHidden->pWC = pWC; 1007 pHidden->pParse = pParse; 1008 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 1009 u16 op; 1010 if( pTerm->leftCursor != pSrc->iCursor ) continue; 1011 if( pTerm->prereqRight & mUnusable ) continue; 1012 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 1013 testcase( pTerm->eOperator & WO_IN ); 1014 testcase( pTerm->eOperator & WO_IS ); 1015 testcase( pTerm->eOperator & WO_ISNULL ); 1016 testcase( pTerm->eOperator & WO_ALL ); 1017 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 1018 if( pTerm->wtFlags & TERM_VNULL ) continue; 1019 1020 /* tag-20191211-002: WHERE-clause constraints are not useful to the 1021 ** right-hand table of a LEFT JOIN. See tag-20191211-001 for the 1022 ** equivalent restriction for ordinary tables. */ 1023 if( (pSrc->fg.jointype & JT_LEFT)!=0 1024 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 1025 ){ 1026 continue; 1027 } 1028 assert( pTerm->u.x.leftColumn>=(-1) ); 1029 pIdxCons[j].iColumn = pTerm->u.x.leftColumn; 1030 pIdxCons[j].iTermOffset = i; 1031 op = pTerm->eOperator & WO_ALL; 1032 if( op==WO_IN ) op = WO_EQ; 1033 if( op==WO_AUX ){ 1034 pIdxCons[j].op = pTerm->eMatchOp; 1035 }else if( op & (WO_ISNULL|WO_IS) ){ 1036 if( op==WO_ISNULL ){ 1037 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_ISNULL; 1038 }else{ 1039 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_IS; 1040 } 1041 }else{ 1042 pIdxCons[j].op = (u8)op; 1043 /* The direct assignment in the previous line is possible only because 1044 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 1045 ** following asserts verify this fact. */ 1046 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 1047 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 1048 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 1049 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 1050 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 1051 assert( pTerm->eOperator&(WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_AUX) ); 1052 1053 if( op & (WO_LT|WO_LE|WO_GT|WO_GE) 1054 && sqlite3ExprIsVector(pTerm->pExpr->pRight) 1055 ){ 1056 testcase( j!=i ); 1057 if( j<16 ) mNoOmit |= (1 << j); 1058 if( op==WO_LT ) pIdxCons[j].op = WO_LE; 1059 if( op==WO_GT ) pIdxCons[j].op = WO_GE; 1060 } 1061 } 1062 1063 j++; 1064 } 1065 pIdxInfo->nConstraint = j; 1066 for(i=0; i<nOrderBy; i++){ 1067 Expr *pExpr = pOrderBy->a[i].pExpr; 1068 pIdxOrderBy[i].iColumn = pExpr->iColumn; 1069 pIdxOrderBy[i].desc = pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC; 1070 } 1071 1072 *pmNoOmit = mNoOmit; 1073 return pIdxInfo; 1074 } 1075 1076 /* 1077 ** The table object reference passed as the second argument to this function 1078 ** must represent a virtual table. This function invokes the xBestIndex() 1079 ** method of the virtual table with the sqlite3_index_info object that 1080 ** comes in as the 3rd argument to this function. 1081 ** 1082 ** If an error occurs, pParse is populated with an error message and an 1083 ** appropriate error code is returned. A return of SQLITE_CONSTRAINT from 1084 ** xBestIndex is not considered an error. SQLITE_CONSTRAINT indicates that 1085 ** the current configuration of "unusable" flags in sqlite3_index_info can 1086 ** not result in a valid plan. 1087 ** 1088 ** Whether or not an error is returned, it is the responsibility of the 1089 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 1090 ** that this is required. 1091 */ 1092 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 1093 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 1094 int rc; 1095 1096 whereTraceIndexInfoInputs(p); 1097 rc = pVtab->pModule->xBestIndex(pVtab, p); 1098 whereTraceIndexInfoOutputs(p); 1099 1100 if( rc!=SQLITE_OK && rc!=SQLITE_CONSTRAINT ){ 1101 if( rc==SQLITE_NOMEM ){ 1102 sqlite3OomFault(pParse->db); 1103 }else if( !pVtab->zErrMsg ){ 1104 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 1105 }else{ 1106 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 1107 } 1108 } 1109 sqlite3_free(pVtab->zErrMsg); 1110 pVtab->zErrMsg = 0; 1111 return rc; 1112 } 1113 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 1114 1115 #ifdef SQLITE_ENABLE_STAT4 1116 /* 1117 ** Estimate the location of a particular key among all keys in an 1118 ** index. Store the results in aStat as follows: 1119 ** 1120 ** aStat[0] Est. number of rows less than pRec 1121 ** aStat[1] Est. number of rows equal to pRec 1122 ** 1123 ** Return the index of the sample that is the smallest sample that 1124 ** is greater than or equal to pRec. Note that this index is not an index 1125 ** into the aSample[] array - it is an index into a virtual set of samples 1126 ** based on the contents of aSample[] and the number of fields in record 1127 ** pRec. 1128 */ 1129 static int whereKeyStats( 1130 Parse *pParse, /* Database connection */ 1131 Index *pIdx, /* Index to consider domain of */ 1132 UnpackedRecord *pRec, /* Vector of values to consider */ 1133 int roundUp, /* Round up if true. Round down if false */ 1134 tRowcnt *aStat /* OUT: stats written here */ 1135 ){ 1136 IndexSample *aSample = pIdx->aSample; 1137 int iCol; /* Index of required stats in anEq[] etc. */ 1138 int i; /* Index of first sample >= pRec */ 1139 int iSample; /* Smallest sample larger than or equal to pRec */ 1140 int iMin = 0; /* Smallest sample not yet tested */ 1141 int iTest; /* Next sample to test */ 1142 int res; /* Result of comparison operation */ 1143 int nField; /* Number of fields in pRec */ 1144 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 1145 1146 #ifndef SQLITE_DEBUG 1147 UNUSED_PARAMETER( pParse ); 1148 #endif 1149 assert( pRec!=0 ); 1150 assert( pIdx->nSample>0 ); 1151 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1152 1153 /* Do a binary search to find the first sample greater than or equal 1154 ** to pRec. If pRec contains a single field, the set of samples to search 1155 ** is simply the aSample[] array. If the samples in aSample[] contain more 1156 ** than one fields, all fields following the first are ignored. 1157 ** 1158 ** If pRec contains N fields, where N is more than one, then as well as the 1159 ** samples in aSample[] (truncated to N fields), the search also has to 1160 ** consider prefixes of those samples. For example, if the set of samples 1161 ** in aSample is: 1162 ** 1163 ** aSample[0] = (a, 5) 1164 ** aSample[1] = (a, 10) 1165 ** aSample[2] = (b, 5) 1166 ** aSample[3] = (c, 100) 1167 ** aSample[4] = (c, 105) 1168 ** 1169 ** Then the search space should ideally be the samples above and the 1170 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1171 ** the code actually searches this set: 1172 ** 1173 ** 0: (a) 1174 ** 1: (a, 5) 1175 ** 2: (a, 10) 1176 ** 3: (a, 10) 1177 ** 4: (b) 1178 ** 5: (b, 5) 1179 ** 6: (c) 1180 ** 7: (c, 100) 1181 ** 8: (c, 105) 1182 ** 9: (c, 105) 1183 ** 1184 ** For each sample in the aSample[] array, N samples are present in the 1185 ** effective sample array. In the above, samples 0 and 1 are based on 1186 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1187 ** 1188 ** Often, sample i of each block of N effective samples has (i+1) fields. 1189 ** Except, each sample may be extended to ensure that it is greater than or 1190 ** equal to the previous sample in the array. For example, in the above, 1191 ** sample 2 is the first sample of a block of N samples, so at first it 1192 ** appears that it should be 1 field in size. However, that would make it 1193 ** smaller than sample 1, so the binary search would not work. As a result, 1194 ** it is extended to two fields. The duplicates that this creates do not 1195 ** cause any problems. 1196 */ 1197 nField = pRec->nField; 1198 iCol = 0; 1199 iSample = pIdx->nSample * nField; 1200 do{ 1201 int iSamp; /* Index in aSample[] of test sample */ 1202 int n; /* Number of fields in test sample */ 1203 1204 iTest = (iMin+iSample)/2; 1205 iSamp = iTest / nField; 1206 if( iSamp>0 ){ 1207 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1208 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1209 ** fields that is greater than the previous effective sample. */ 1210 for(n=(iTest % nField) + 1; n<nField; n++){ 1211 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1212 } 1213 }else{ 1214 n = iTest + 1; 1215 } 1216 1217 pRec->nField = n; 1218 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1219 if( res<0 ){ 1220 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1221 iMin = iTest+1; 1222 }else if( res==0 && n<nField ){ 1223 iLower = aSample[iSamp].anLt[n-1]; 1224 iMin = iTest+1; 1225 res = -1; 1226 }else{ 1227 iSample = iTest; 1228 iCol = n-1; 1229 } 1230 }while( res && iMin<iSample ); 1231 i = iSample / nField; 1232 1233 #ifdef SQLITE_DEBUG 1234 /* The following assert statements check that the binary search code 1235 ** above found the right answer. This block serves no purpose other 1236 ** than to invoke the asserts. */ 1237 if( pParse->db->mallocFailed==0 ){ 1238 if( res==0 ){ 1239 /* If (res==0) is true, then pRec must be equal to sample i. */ 1240 assert( i<pIdx->nSample ); 1241 assert( iCol==nField-1 ); 1242 pRec->nField = nField; 1243 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1244 || pParse->db->mallocFailed 1245 ); 1246 }else{ 1247 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1248 ** all samples in the aSample[] array, pRec must be smaller than the 1249 ** (iCol+1) field prefix of sample i. */ 1250 assert( i<=pIdx->nSample && i>=0 ); 1251 pRec->nField = iCol+1; 1252 assert( i==pIdx->nSample 1253 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1254 || pParse->db->mallocFailed ); 1255 1256 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1257 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1258 ** be greater than or equal to the (iCol) field prefix of sample i. 1259 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1260 if( iCol>0 ){ 1261 pRec->nField = iCol; 1262 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1263 || pParse->db->mallocFailed ); 1264 } 1265 if( i>0 ){ 1266 pRec->nField = nField; 1267 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1268 || pParse->db->mallocFailed ); 1269 } 1270 } 1271 } 1272 #endif /* ifdef SQLITE_DEBUG */ 1273 1274 if( res==0 ){ 1275 /* Record pRec is equal to sample i */ 1276 assert( iCol==nField-1 ); 1277 aStat[0] = aSample[i].anLt[iCol]; 1278 aStat[1] = aSample[i].anEq[iCol]; 1279 }else{ 1280 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1281 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1282 ** is larger than all samples in the array. */ 1283 tRowcnt iUpper, iGap; 1284 if( i>=pIdx->nSample ){ 1285 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1286 }else{ 1287 iUpper = aSample[i].anLt[iCol]; 1288 } 1289 1290 if( iLower>=iUpper ){ 1291 iGap = 0; 1292 }else{ 1293 iGap = iUpper - iLower; 1294 } 1295 if( roundUp ){ 1296 iGap = (iGap*2)/3; 1297 }else{ 1298 iGap = iGap/3; 1299 } 1300 aStat[0] = iLower + iGap; 1301 aStat[1] = pIdx->aAvgEq[nField-1]; 1302 } 1303 1304 /* Restore the pRec->nField value before returning. */ 1305 pRec->nField = nField; 1306 return i; 1307 } 1308 #endif /* SQLITE_ENABLE_STAT4 */ 1309 1310 /* 1311 ** If it is not NULL, pTerm is a term that provides an upper or lower 1312 ** bound on a range scan. Without considering pTerm, it is estimated 1313 ** that the scan will visit nNew rows. This function returns the number 1314 ** estimated to be visited after taking pTerm into account. 1315 ** 1316 ** If the user explicitly specified a likelihood() value for this term, 1317 ** then the return value is the likelihood multiplied by the number of 1318 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1319 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1320 */ 1321 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1322 LogEst nRet = nNew; 1323 if( pTerm ){ 1324 if( pTerm->truthProb<=0 ){ 1325 nRet += pTerm->truthProb; 1326 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1327 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1328 } 1329 } 1330 return nRet; 1331 } 1332 1333 1334 #ifdef SQLITE_ENABLE_STAT4 1335 /* 1336 ** Return the affinity for a single column of an index. 1337 */ 1338 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1339 assert( iCol>=0 && iCol<pIdx->nColumn ); 1340 if( !pIdx->zColAff ){ 1341 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1342 } 1343 assert( pIdx->zColAff[iCol]!=0 ); 1344 return pIdx->zColAff[iCol]; 1345 } 1346 #endif 1347 1348 1349 #ifdef SQLITE_ENABLE_STAT4 1350 /* 1351 ** This function is called to estimate the number of rows visited by a 1352 ** range-scan on a skip-scan index. For example: 1353 ** 1354 ** CREATE INDEX i1 ON t1(a, b, c); 1355 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1356 ** 1357 ** Value pLoop->nOut is currently set to the estimated number of rows 1358 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1359 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1360 ** on the stat4 data for the index. this scan will be peformed multiple 1361 ** times (once for each (a,b) combination that matches a=?) is dealt with 1362 ** by the caller. 1363 ** 1364 ** It does this by scanning through all stat4 samples, comparing values 1365 ** extracted from pLower and pUpper with the corresponding column in each 1366 ** sample. If L and U are the number of samples found to be less than or 1367 ** equal to the values extracted from pLower and pUpper respectively, and 1368 ** N is the total number of samples, the pLoop->nOut value is adjusted 1369 ** as follows: 1370 ** 1371 ** nOut = nOut * ( min(U - L, 1) / N ) 1372 ** 1373 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1374 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1375 ** U is set to N. 1376 ** 1377 ** Normally, this function sets *pbDone to 1 before returning. However, 1378 ** if no value can be extracted from either pLower or pUpper (and so the 1379 ** estimate of the number of rows delivered remains unchanged), *pbDone 1380 ** is left as is. 1381 ** 1382 ** If an error occurs, an SQLite error code is returned. Otherwise, 1383 ** SQLITE_OK. 1384 */ 1385 static int whereRangeSkipScanEst( 1386 Parse *pParse, /* Parsing & code generating context */ 1387 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1388 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1389 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1390 int *pbDone /* Set to true if at least one expr. value extracted */ 1391 ){ 1392 Index *p = pLoop->u.btree.pIndex; 1393 int nEq = pLoop->u.btree.nEq; 1394 sqlite3 *db = pParse->db; 1395 int nLower = -1; 1396 int nUpper = p->nSample+1; 1397 int rc = SQLITE_OK; 1398 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1399 CollSeq *pColl; 1400 1401 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1402 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1403 sqlite3_value *pVal = 0; /* Value extracted from record */ 1404 1405 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1406 if( pLower ){ 1407 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1408 nLower = 0; 1409 } 1410 if( pUpper && rc==SQLITE_OK ){ 1411 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1412 nUpper = p2 ? 0 : p->nSample; 1413 } 1414 1415 if( p1 || p2 ){ 1416 int i; 1417 int nDiff; 1418 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1419 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1420 if( rc==SQLITE_OK && p1 ){ 1421 int res = sqlite3MemCompare(p1, pVal, pColl); 1422 if( res>=0 ) nLower++; 1423 } 1424 if( rc==SQLITE_OK && p2 ){ 1425 int res = sqlite3MemCompare(p2, pVal, pColl); 1426 if( res>=0 ) nUpper++; 1427 } 1428 } 1429 nDiff = (nUpper - nLower); 1430 if( nDiff<=0 ) nDiff = 1; 1431 1432 /* If there is both an upper and lower bound specified, and the 1433 ** comparisons indicate that they are close together, use the fallback 1434 ** method (assume that the scan visits 1/64 of the rows) for estimating 1435 ** the number of rows visited. Otherwise, estimate the number of rows 1436 ** using the method described in the header comment for this function. */ 1437 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1438 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1439 pLoop->nOut -= nAdjust; 1440 *pbDone = 1; 1441 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1442 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1443 } 1444 1445 }else{ 1446 assert( *pbDone==0 ); 1447 } 1448 1449 sqlite3ValueFree(p1); 1450 sqlite3ValueFree(p2); 1451 sqlite3ValueFree(pVal); 1452 1453 return rc; 1454 } 1455 #endif /* SQLITE_ENABLE_STAT4 */ 1456 1457 /* 1458 ** This function is used to estimate the number of rows that will be visited 1459 ** by scanning an index for a range of values. The range may have an upper 1460 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1461 ** and lower bounds are represented by pLower and pUpper respectively. For 1462 ** example, assuming that index p is on t1(a): 1463 ** 1464 ** ... FROM t1 WHERE a > ? AND a < ? ... 1465 ** |_____| |_____| 1466 ** | | 1467 ** pLower pUpper 1468 ** 1469 ** If either of the upper or lower bound is not present, then NULL is passed in 1470 ** place of the corresponding WhereTerm. 1471 ** 1472 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1473 ** column subject to the range constraint. Or, equivalently, the number of 1474 ** equality constraints optimized by the proposed index scan. For example, 1475 ** assuming index p is on t1(a, b), and the SQL query is: 1476 ** 1477 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1478 ** 1479 ** then nEq is set to 1 (as the range restricted column, b, is the second 1480 ** left-most column of the index). Or, if the query is: 1481 ** 1482 ** ... FROM t1 WHERE a > ? AND a < ? ... 1483 ** 1484 ** then nEq is set to 0. 1485 ** 1486 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1487 ** number of rows that the index scan is expected to visit without 1488 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1489 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1490 ** to account for the range constraints pLower and pUpper. 1491 ** 1492 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1493 ** used, a single range inequality reduces the search space by a factor of 4. 1494 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1495 ** rows visited by a factor of 64. 1496 */ 1497 static int whereRangeScanEst( 1498 Parse *pParse, /* Parsing & code generating context */ 1499 WhereLoopBuilder *pBuilder, 1500 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1501 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1502 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1503 ){ 1504 int rc = SQLITE_OK; 1505 int nOut = pLoop->nOut; 1506 LogEst nNew; 1507 1508 #ifdef SQLITE_ENABLE_STAT4 1509 Index *p = pLoop->u.btree.pIndex; 1510 int nEq = pLoop->u.btree.nEq; 1511 1512 if( p->nSample>0 && ALWAYS(nEq<p->nSampleCol) 1513 && OptimizationEnabled(pParse->db, SQLITE_Stat4) 1514 ){ 1515 if( nEq==pBuilder->nRecValid ){ 1516 UnpackedRecord *pRec = pBuilder->pRec; 1517 tRowcnt a[2]; 1518 int nBtm = pLoop->u.btree.nBtm; 1519 int nTop = pLoop->u.btree.nTop; 1520 1521 /* Variable iLower will be set to the estimate of the number of rows in 1522 ** the index that are less than the lower bound of the range query. The 1523 ** lower bound being the concatenation of $P and $L, where $P is the 1524 ** key-prefix formed by the nEq values matched against the nEq left-most 1525 ** columns of the index, and $L is the value in pLower. 1526 ** 1527 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1528 ** is not a simple variable or literal value), the lower bound of the 1529 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1530 ** if $L is available, whereKeyStats() is called for both ($P) and 1531 ** ($P:$L) and the larger of the two returned values is used. 1532 ** 1533 ** Similarly, iUpper is to be set to the estimate of the number of rows 1534 ** less than the upper bound of the range query. Where the upper bound 1535 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1536 ** of iUpper are requested of whereKeyStats() and the smaller used. 1537 ** 1538 ** The number of rows between the two bounds is then just iUpper-iLower. 1539 */ 1540 tRowcnt iLower; /* Rows less than the lower bound */ 1541 tRowcnt iUpper; /* Rows less than the upper bound */ 1542 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1543 int iUprIdx = -1; /* aSample[] for the upper bound */ 1544 1545 if( pRec ){ 1546 testcase( pRec->nField!=pBuilder->nRecValid ); 1547 pRec->nField = pBuilder->nRecValid; 1548 } 1549 /* Determine iLower and iUpper using ($P) only. */ 1550 if( nEq==0 ){ 1551 iLower = 0; 1552 iUpper = p->nRowEst0; 1553 }else{ 1554 /* Note: this call could be optimized away - since the same values must 1555 ** have been requested when testing key $P in whereEqualScanEst(). */ 1556 whereKeyStats(pParse, p, pRec, 0, a); 1557 iLower = a[0]; 1558 iUpper = a[0] + a[1]; 1559 } 1560 1561 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1562 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1563 assert( p->aSortOrder!=0 ); 1564 if( p->aSortOrder[nEq] ){ 1565 /* The roles of pLower and pUpper are swapped for a DESC index */ 1566 SWAP(WhereTerm*, pLower, pUpper); 1567 SWAP(int, nBtm, nTop); 1568 } 1569 1570 /* If possible, improve on the iLower estimate using ($P:$L). */ 1571 if( pLower ){ 1572 int n; /* Values extracted from pExpr */ 1573 Expr *pExpr = pLower->pExpr->pRight; 1574 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1575 if( rc==SQLITE_OK && n ){ 1576 tRowcnt iNew; 1577 u16 mask = WO_GT|WO_LE; 1578 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1579 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1580 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1581 if( iNew>iLower ) iLower = iNew; 1582 nOut--; 1583 pLower = 0; 1584 } 1585 } 1586 1587 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1588 if( pUpper ){ 1589 int n; /* Values extracted from pExpr */ 1590 Expr *pExpr = pUpper->pExpr->pRight; 1591 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1592 if( rc==SQLITE_OK && n ){ 1593 tRowcnt iNew; 1594 u16 mask = WO_GT|WO_LE; 1595 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1596 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1597 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1598 if( iNew<iUpper ) iUpper = iNew; 1599 nOut--; 1600 pUpper = 0; 1601 } 1602 } 1603 1604 pBuilder->pRec = pRec; 1605 if( rc==SQLITE_OK ){ 1606 if( iUpper>iLower ){ 1607 nNew = sqlite3LogEst(iUpper - iLower); 1608 /* TUNING: If both iUpper and iLower are derived from the same 1609 ** sample, then assume they are 4x more selective. This brings 1610 ** the estimated selectivity more in line with what it would be 1611 ** if estimated without the use of STAT4 tables. */ 1612 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1613 }else{ 1614 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1615 } 1616 if( nNew<nOut ){ 1617 nOut = nNew; 1618 } 1619 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1620 (u32)iLower, (u32)iUpper, nOut)); 1621 } 1622 }else{ 1623 int bDone = 0; 1624 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1625 if( bDone ) return rc; 1626 } 1627 } 1628 #else 1629 UNUSED_PARAMETER(pParse); 1630 UNUSED_PARAMETER(pBuilder); 1631 assert( pLower || pUpper ); 1632 #endif 1633 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1634 nNew = whereRangeAdjust(pLower, nOut); 1635 nNew = whereRangeAdjust(pUpper, nNew); 1636 1637 /* TUNING: If there is both an upper and lower limit and neither limit 1638 ** has an application-defined likelihood(), assume the range is 1639 ** reduced by an additional 75%. This means that, by default, an open-ended 1640 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1641 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1642 ** match 1/64 of the index. */ 1643 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1644 nNew -= 20; 1645 } 1646 1647 nOut -= (pLower!=0) + (pUpper!=0); 1648 if( nNew<10 ) nNew = 10; 1649 if( nNew<nOut ) nOut = nNew; 1650 #if defined(WHERETRACE_ENABLED) 1651 if( pLoop->nOut>nOut ){ 1652 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1653 pLoop->nOut, nOut)); 1654 } 1655 #endif 1656 pLoop->nOut = (LogEst)nOut; 1657 return rc; 1658 } 1659 1660 #ifdef SQLITE_ENABLE_STAT4 1661 /* 1662 ** Estimate the number of rows that will be returned based on 1663 ** an equality constraint x=VALUE and where that VALUE occurs in 1664 ** the histogram data. This only works when x is the left-most 1665 ** column of an index and sqlite_stat4 histogram data is available 1666 ** for that index. When pExpr==NULL that means the constraint is 1667 ** "x IS NULL" instead of "x=VALUE". 1668 ** 1669 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1670 ** If unable to make an estimate, leave *pnRow unchanged and return 1671 ** non-zero. 1672 ** 1673 ** This routine can fail if it is unable to load a collating sequence 1674 ** required for string comparison, or if unable to allocate memory 1675 ** for a UTF conversion required for comparison. The error is stored 1676 ** in the pParse structure. 1677 */ 1678 static int whereEqualScanEst( 1679 Parse *pParse, /* Parsing & code generating context */ 1680 WhereLoopBuilder *pBuilder, 1681 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1682 tRowcnt *pnRow /* Write the revised row estimate here */ 1683 ){ 1684 Index *p = pBuilder->pNew->u.btree.pIndex; 1685 int nEq = pBuilder->pNew->u.btree.nEq; 1686 UnpackedRecord *pRec = pBuilder->pRec; 1687 int rc; /* Subfunction return code */ 1688 tRowcnt a[2]; /* Statistics */ 1689 int bOk; 1690 1691 assert( nEq>=1 ); 1692 assert( nEq<=p->nColumn ); 1693 assert( p->aSample!=0 ); 1694 assert( p->nSample>0 ); 1695 assert( pBuilder->nRecValid<nEq ); 1696 1697 /* If values are not available for all fields of the index to the left 1698 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1699 if( pBuilder->nRecValid<(nEq-1) ){ 1700 return SQLITE_NOTFOUND; 1701 } 1702 1703 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1704 ** below would return the same value. */ 1705 if( nEq>=p->nColumn ){ 1706 *pnRow = 1; 1707 return SQLITE_OK; 1708 } 1709 1710 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1711 pBuilder->pRec = pRec; 1712 if( rc!=SQLITE_OK ) return rc; 1713 if( bOk==0 ) return SQLITE_NOTFOUND; 1714 pBuilder->nRecValid = nEq; 1715 1716 whereKeyStats(pParse, p, pRec, 0, a); 1717 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1718 p->zName, nEq-1, (int)a[1])); 1719 *pnRow = a[1]; 1720 1721 return rc; 1722 } 1723 #endif /* SQLITE_ENABLE_STAT4 */ 1724 1725 #ifdef SQLITE_ENABLE_STAT4 1726 /* 1727 ** Estimate the number of rows that will be returned based on 1728 ** an IN constraint where the right-hand side of the IN operator 1729 ** is a list of values. Example: 1730 ** 1731 ** WHERE x IN (1,2,3,4) 1732 ** 1733 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1734 ** If unable to make an estimate, leave *pnRow unchanged and return 1735 ** non-zero. 1736 ** 1737 ** This routine can fail if it is unable to load a collating sequence 1738 ** required for string comparison, or if unable to allocate memory 1739 ** for a UTF conversion required for comparison. The error is stored 1740 ** in the pParse structure. 1741 */ 1742 static int whereInScanEst( 1743 Parse *pParse, /* Parsing & code generating context */ 1744 WhereLoopBuilder *pBuilder, 1745 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1746 tRowcnt *pnRow /* Write the revised row estimate here */ 1747 ){ 1748 Index *p = pBuilder->pNew->u.btree.pIndex; 1749 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1750 int nRecValid = pBuilder->nRecValid; 1751 int rc = SQLITE_OK; /* Subfunction return code */ 1752 tRowcnt nEst; /* Number of rows for a single term */ 1753 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1754 int i; /* Loop counter */ 1755 1756 assert( p->aSample!=0 ); 1757 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1758 nEst = nRow0; 1759 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1760 nRowEst += nEst; 1761 pBuilder->nRecValid = nRecValid; 1762 } 1763 1764 if( rc==SQLITE_OK ){ 1765 if( nRowEst > nRow0 ) nRowEst = nRow0; 1766 *pnRow = nRowEst; 1767 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1768 } 1769 assert( pBuilder->nRecValid==nRecValid ); 1770 return rc; 1771 } 1772 #endif /* SQLITE_ENABLE_STAT4 */ 1773 1774 1775 #ifdef WHERETRACE_ENABLED 1776 /* 1777 ** Print the content of a WhereTerm object 1778 */ 1779 void sqlite3WhereTermPrint(WhereTerm *pTerm, int iTerm){ 1780 if( pTerm==0 ){ 1781 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1782 }else{ 1783 char zType[8]; 1784 char zLeft[50]; 1785 memcpy(zType, "....", 5); 1786 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1787 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1788 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1789 if( pTerm->wtFlags & TERM_CODED ) zType[3] = 'C'; 1790 if( pTerm->eOperator & WO_SINGLE ){ 1791 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 1792 pTerm->leftCursor, pTerm->u.x.leftColumn); 1793 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 1794 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%llx", 1795 pTerm->u.pOrInfo->indexable); 1796 }else{ 1797 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 1798 } 1799 sqlite3DebugPrintf( 1800 "TERM-%-3d %p %s %-12s op=%03x wtFlags=%04x", 1801 iTerm, pTerm, zType, zLeft, pTerm->eOperator, pTerm->wtFlags); 1802 /* The 0x10000 .wheretrace flag causes extra information to be 1803 ** shown about each Term */ 1804 if( sqlite3WhereTrace & 0x10000 ){ 1805 sqlite3DebugPrintf(" prob=%-3d prereq=%llx,%llx", 1806 pTerm->truthProb, (u64)pTerm->prereqAll, (u64)pTerm->prereqRight); 1807 } 1808 if( pTerm->u.x.iField ){ 1809 sqlite3DebugPrintf(" iField=%d", pTerm->u.x.iField); 1810 } 1811 if( pTerm->iParent>=0 ){ 1812 sqlite3DebugPrintf(" iParent=%d", pTerm->iParent); 1813 } 1814 sqlite3DebugPrintf("\n"); 1815 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 1816 } 1817 } 1818 #endif 1819 1820 #ifdef WHERETRACE_ENABLED 1821 /* 1822 ** Show the complete content of a WhereClause 1823 */ 1824 void sqlite3WhereClausePrint(WhereClause *pWC){ 1825 int i; 1826 for(i=0; i<pWC->nTerm; i++){ 1827 sqlite3WhereTermPrint(&pWC->a[i], i); 1828 } 1829 } 1830 #endif 1831 1832 #ifdef WHERETRACE_ENABLED 1833 /* 1834 ** Print a WhereLoop object for debugging purposes 1835 */ 1836 void sqlite3WhereLoopPrint(WhereLoop *p, WhereClause *pWC){ 1837 WhereInfo *pWInfo = pWC->pWInfo; 1838 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 1839 SrcItem *pItem = pWInfo->pTabList->a + p->iTab; 1840 Table *pTab = pItem->pTab; 1841 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 1842 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 1843 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 1844 sqlite3DebugPrintf(" %12s", 1845 pItem->zAlias ? pItem->zAlias : pTab->zName); 1846 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 1847 const char *zName; 1848 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 1849 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 1850 int i = sqlite3Strlen30(zName) - 1; 1851 while( zName[i]!='_' ) i--; 1852 zName += i; 1853 } 1854 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 1855 }else{ 1856 sqlite3DebugPrintf("%20s",""); 1857 } 1858 }else{ 1859 char *z; 1860 if( p->u.vtab.idxStr ){ 1861 z = sqlite3_mprintf("(%d,\"%s\",%#x)", 1862 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 1863 }else{ 1864 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 1865 } 1866 sqlite3DebugPrintf(" %-19s", z); 1867 sqlite3_free(z); 1868 } 1869 if( p->wsFlags & WHERE_SKIPSCAN ){ 1870 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 1871 }else{ 1872 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); 1873 } 1874 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 1875 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 1876 int i; 1877 for(i=0; i<p->nLTerm; i++){ 1878 sqlite3WhereTermPrint(p->aLTerm[i], i); 1879 } 1880 } 1881 } 1882 #endif 1883 1884 /* 1885 ** Convert bulk memory into a valid WhereLoop that can be passed 1886 ** to whereLoopClear harmlessly. 1887 */ 1888 static void whereLoopInit(WhereLoop *p){ 1889 p->aLTerm = p->aLTermSpace; 1890 p->nLTerm = 0; 1891 p->nLSlot = ArraySize(p->aLTermSpace); 1892 p->wsFlags = 0; 1893 } 1894 1895 /* 1896 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 1897 */ 1898 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 1899 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 1900 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 1901 sqlite3_free(p->u.vtab.idxStr); 1902 p->u.vtab.needFree = 0; 1903 p->u.vtab.idxStr = 0; 1904 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 1905 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 1906 sqlite3DbFreeNN(db, p->u.btree.pIndex); 1907 p->u.btree.pIndex = 0; 1908 } 1909 } 1910 } 1911 1912 /* 1913 ** Deallocate internal memory used by a WhereLoop object 1914 */ 1915 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 1916 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1917 whereLoopClearUnion(db, p); 1918 whereLoopInit(p); 1919 } 1920 1921 /* 1922 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 1923 */ 1924 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 1925 WhereTerm **paNew; 1926 if( p->nLSlot>=n ) return SQLITE_OK; 1927 n = (n+7)&~7; 1928 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 1929 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 1930 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 1931 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1932 p->aLTerm = paNew; 1933 p->nLSlot = n; 1934 return SQLITE_OK; 1935 } 1936 1937 /* 1938 ** Transfer content from the second pLoop into the first. 1939 */ 1940 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 1941 whereLoopClearUnion(db, pTo); 1942 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 1943 memset(&pTo->u, 0, sizeof(pTo->u)); 1944 return SQLITE_NOMEM_BKPT; 1945 } 1946 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 1947 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 1948 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 1949 pFrom->u.vtab.needFree = 0; 1950 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 1951 pFrom->u.btree.pIndex = 0; 1952 } 1953 return SQLITE_OK; 1954 } 1955 1956 /* 1957 ** Delete a WhereLoop object 1958 */ 1959 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 1960 whereLoopClear(db, p); 1961 sqlite3DbFreeNN(db, p); 1962 } 1963 1964 /* 1965 ** Free a WhereInfo structure 1966 */ 1967 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 1968 int i; 1969 assert( pWInfo!=0 ); 1970 for(i=0; i<pWInfo->nLevel; i++){ 1971 WhereLevel *pLevel = &pWInfo->a[i]; 1972 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){ 1973 sqlite3DbFree(db, pLevel->u.in.aInLoop); 1974 } 1975 } 1976 sqlite3WhereClauseClear(&pWInfo->sWC); 1977 while( pWInfo->pLoops ){ 1978 WhereLoop *p = pWInfo->pLoops; 1979 pWInfo->pLoops = p->pNextLoop; 1980 whereLoopDelete(db, p); 1981 } 1982 assert( pWInfo->pExprMods==0 ); 1983 sqlite3DbFreeNN(db, pWInfo); 1984 } 1985 1986 /* 1987 ** Return TRUE if all of the following are true: 1988 ** 1989 ** (1) X has the same or lower cost that Y 1990 ** (2) X uses fewer WHERE clause terms than Y 1991 ** (3) Every WHERE clause term used by X is also used by Y 1992 ** (4) X skips at least as many columns as Y 1993 ** (5) If X is a covering index, than Y is too 1994 ** 1995 ** Conditions (2) and (3) mean that X is a "proper subset" of Y. 1996 ** If X is a proper subset of Y then Y is a better choice and ought 1997 ** to have a lower cost. This routine returns TRUE when that cost 1998 ** relationship is inverted and needs to be adjusted. Constraint (4) 1999 ** was added because if X uses skip-scan less than Y it still might 2000 ** deserve a lower cost even if it is a proper subset of Y. Constraint (5) 2001 ** was added because a covering index probably deserves to have a lower cost 2002 ** than a non-covering index even if it is a proper subset. 2003 */ 2004 static int whereLoopCheaperProperSubset( 2005 const WhereLoop *pX, /* First WhereLoop to compare */ 2006 const WhereLoop *pY /* Compare against this WhereLoop */ 2007 ){ 2008 int i, j; 2009 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 2010 return 0; /* X is not a subset of Y */ 2011 } 2012 if( pY->nSkip > pX->nSkip ) return 0; 2013 if( pX->rRun >= pY->rRun ){ 2014 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ 2015 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ 2016 } 2017 for(i=pX->nLTerm-1; i>=0; i--){ 2018 if( pX->aLTerm[i]==0 ) continue; 2019 for(j=pY->nLTerm-1; j>=0; j--){ 2020 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 2021 } 2022 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 2023 } 2024 if( (pX->wsFlags&WHERE_IDX_ONLY)!=0 2025 && (pY->wsFlags&WHERE_IDX_ONLY)==0 ){ 2026 return 0; /* Constraint (5) */ 2027 } 2028 return 1; /* All conditions meet */ 2029 } 2030 2031 /* 2032 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so 2033 ** that: 2034 ** 2035 ** (1) pTemplate costs less than any other WhereLoops that are a proper 2036 ** subset of pTemplate 2037 ** 2038 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 2039 ** is a proper subset. 2040 ** 2041 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 2042 ** WHERE clause terms than Y and that every WHERE clause term used by X is 2043 ** also used by Y. 2044 */ 2045 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 2046 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 2047 for(; p; p=p->pNextLoop){ 2048 if( p->iTab!=pTemplate->iTab ) continue; 2049 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 2050 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 2051 /* Adjust pTemplate cost downward so that it is cheaper than its 2052 ** subset p. */ 2053 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2054 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); 2055 pTemplate->rRun = p->rRun; 2056 pTemplate->nOut = p->nOut - 1; 2057 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 2058 /* Adjust pTemplate cost upward so that it is costlier than p since 2059 ** pTemplate is a proper subset of p */ 2060 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2061 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); 2062 pTemplate->rRun = p->rRun; 2063 pTemplate->nOut = p->nOut + 1; 2064 } 2065 } 2066 } 2067 2068 /* 2069 ** Search the list of WhereLoops in *ppPrev looking for one that can be 2070 ** replaced by pTemplate. 2071 ** 2072 ** Return NULL if pTemplate does not belong on the WhereLoop list. 2073 ** In other words if pTemplate ought to be dropped from further consideration. 2074 ** 2075 ** If pX is a WhereLoop that pTemplate can replace, then return the 2076 ** link that points to pX. 2077 ** 2078 ** If pTemplate cannot replace any existing element of the list but needs 2079 ** to be added to the list as a new entry, then return a pointer to the 2080 ** tail of the list. 2081 */ 2082 static WhereLoop **whereLoopFindLesser( 2083 WhereLoop **ppPrev, 2084 const WhereLoop *pTemplate 2085 ){ 2086 WhereLoop *p; 2087 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 2088 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 2089 /* If either the iTab or iSortIdx values for two WhereLoop are different 2090 ** then those WhereLoops need to be considered separately. Neither is 2091 ** a candidate to replace the other. */ 2092 continue; 2093 } 2094 /* In the current implementation, the rSetup value is either zero 2095 ** or the cost of building an automatic index (NlogN) and the NlogN 2096 ** is the same for compatible WhereLoops. */ 2097 assert( p->rSetup==0 || pTemplate->rSetup==0 2098 || p->rSetup==pTemplate->rSetup ); 2099 2100 /* whereLoopAddBtree() always generates and inserts the automatic index 2101 ** case first. Hence compatible candidate WhereLoops never have a larger 2102 ** rSetup. Call this SETUP-INVARIANT */ 2103 assert( p->rSetup>=pTemplate->rSetup ); 2104 2105 /* Any loop using an appliation-defined index (or PRIMARY KEY or 2106 ** UNIQUE constraint) with one or more == constraints is better 2107 ** than an automatic index. Unless it is a skip-scan. */ 2108 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 2109 && (pTemplate->nSkip)==0 2110 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 2111 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 2112 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2113 ){ 2114 break; 2115 } 2116 2117 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2118 ** discarded. WhereLoop p is better if: 2119 ** (1) p has no more dependencies than pTemplate, and 2120 ** (2) p has an equal or lower cost than pTemplate 2121 */ 2122 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2123 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2124 && p->rRun<=pTemplate->rRun /* (2b) */ 2125 && p->nOut<=pTemplate->nOut /* (2c) */ 2126 ){ 2127 return 0; /* Discard pTemplate */ 2128 } 2129 2130 /* If pTemplate is always better than p, then cause p to be overwritten 2131 ** with pTemplate. pTemplate is better than p if: 2132 ** (1) pTemplate has no more dependences than p, and 2133 ** (2) pTemplate has an equal or lower cost than p. 2134 */ 2135 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2136 && p->rRun>=pTemplate->rRun /* (2a) */ 2137 && p->nOut>=pTemplate->nOut /* (2b) */ 2138 ){ 2139 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2140 break; /* Cause p to be overwritten by pTemplate */ 2141 } 2142 } 2143 return ppPrev; 2144 } 2145 2146 /* 2147 ** Insert or replace a WhereLoop entry using the template supplied. 2148 ** 2149 ** An existing WhereLoop entry might be overwritten if the new template 2150 ** is better and has fewer dependencies. Or the template will be ignored 2151 ** and no insert will occur if an existing WhereLoop is faster and has 2152 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2153 ** added based on the template. 2154 ** 2155 ** If pBuilder->pOrSet is not NULL then we care about only the 2156 ** prerequisites and rRun and nOut costs of the N best loops. That 2157 ** information is gathered in the pBuilder->pOrSet object. This special 2158 ** processing mode is used only for OR clause processing. 2159 ** 2160 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2161 ** still might overwrite similar loops with the new template if the 2162 ** new template is better. Loops may be overwritten if the following 2163 ** conditions are met: 2164 ** 2165 ** (1) They have the same iTab. 2166 ** (2) They have the same iSortIdx. 2167 ** (3) The template has same or fewer dependencies than the current loop 2168 ** (4) The template has the same or lower cost than the current loop 2169 */ 2170 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2171 WhereLoop **ppPrev, *p; 2172 WhereInfo *pWInfo = pBuilder->pWInfo; 2173 sqlite3 *db = pWInfo->pParse->db; 2174 int rc; 2175 2176 /* Stop the search once we hit the query planner search limit */ 2177 if( pBuilder->iPlanLimit==0 ){ 2178 WHERETRACE(0xffffffff,("=== query planner search limit reached ===\n")); 2179 if( pBuilder->pOrSet ) pBuilder->pOrSet->n = 0; 2180 return SQLITE_DONE; 2181 } 2182 pBuilder->iPlanLimit--; 2183 2184 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2185 2186 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2187 ** and prereqs. 2188 */ 2189 if( pBuilder->pOrSet!=0 ){ 2190 if( pTemplate->nLTerm ){ 2191 #if WHERETRACE_ENABLED 2192 u16 n = pBuilder->pOrSet->n; 2193 int x = 2194 #endif 2195 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2196 pTemplate->nOut); 2197 #if WHERETRACE_ENABLED /* 0x8 */ 2198 if( sqlite3WhereTrace & 0x8 ){ 2199 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2200 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2201 } 2202 #endif 2203 } 2204 return SQLITE_OK; 2205 } 2206 2207 /* Look for an existing WhereLoop to replace with pTemplate 2208 */ 2209 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2210 2211 if( ppPrev==0 ){ 2212 /* There already exists a WhereLoop on the list that is better 2213 ** than pTemplate, so just ignore pTemplate */ 2214 #if WHERETRACE_ENABLED /* 0x8 */ 2215 if( sqlite3WhereTrace & 0x8 ){ 2216 sqlite3DebugPrintf(" skip: "); 2217 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2218 } 2219 #endif 2220 return SQLITE_OK; 2221 }else{ 2222 p = *ppPrev; 2223 } 2224 2225 /* If we reach this point it means that either p[] should be overwritten 2226 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2227 ** WhereLoop and insert it. 2228 */ 2229 #if WHERETRACE_ENABLED /* 0x8 */ 2230 if( sqlite3WhereTrace & 0x8 ){ 2231 if( p!=0 ){ 2232 sqlite3DebugPrintf("replace: "); 2233 sqlite3WhereLoopPrint(p, pBuilder->pWC); 2234 sqlite3DebugPrintf(" with: "); 2235 }else{ 2236 sqlite3DebugPrintf(" add: "); 2237 } 2238 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2239 } 2240 #endif 2241 if( p==0 ){ 2242 /* Allocate a new WhereLoop to add to the end of the list */ 2243 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2244 if( p==0 ) return SQLITE_NOMEM_BKPT; 2245 whereLoopInit(p); 2246 p->pNextLoop = 0; 2247 }else{ 2248 /* We will be overwriting WhereLoop p[]. But before we do, first 2249 ** go through the rest of the list and delete any other entries besides 2250 ** p[] that are also supplated by pTemplate */ 2251 WhereLoop **ppTail = &p->pNextLoop; 2252 WhereLoop *pToDel; 2253 while( *ppTail ){ 2254 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2255 if( ppTail==0 ) break; 2256 pToDel = *ppTail; 2257 if( pToDel==0 ) break; 2258 *ppTail = pToDel->pNextLoop; 2259 #if WHERETRACE_ENABLED /* 0x8 */ 2260 if( sqlite3WhereTrace & 0x8 ){ 2261 sqlite3DebugPrintf(" delete: "); 2262 sqlite3WhereLoopPrint(pToDel, pBuilder->pWC); 2263 } 2264 #endif 2265 whereLoopDelete(db, pToDel); 2266 } 2267 } 2268 rc = whereLoopXfer(db, p, pTemplate); 2269 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2270 Index *pIndex = p->u.btree.pIndex; 2271 if( pIndex && pIndex->idxType==SQLITE_IDXTYPE_IPK ){ 2272 p->u.btree.pIndex = 0; 2273 } 2274 } 2275 return rc; 2276 } 2277 2278 /* 2279 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2280 ** WHERE clause that reference the loop but which are not used by an 2281 ** index. 2282 * 2283 ** For every WHERE clause term that is not used by the index 2284 ** and which has a truth probability assigned by one of the likelihood(), 2285 ** likely(), or unlikely() SQL functions, reduce the estimated number 2286 ** of output rows by the probability specified. 2287 ** 2288 ** TUNING: For every WHERE clause term that is not used by the index 2289 ** and which does not have an assigned truth probability, heuristics 2290 ** described below are used to try to estimate the truth probability. 2291 ** TODO --> Perhaps this is something that could be improved by better 2292 ** table statistics. 2293 ** 2294 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2295 ** value corresponds to -1 in LogEst notation, so this means decrement 2296 ** the WhereLoop.nOut field for every such WHERE clause term. 2297 ** 2298 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2299 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2300 ** final output row estimate is no greater than 1/4 of the total number 2301 ** of rows in the table. In other words, assume that x==EXPR will filter 2302 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2303 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2304 ** on the "x" column and so in that case only cap the output row estimate 2305 ** at 1/2 instead of 1/4. 2306 */ 2307 static void whereLoopOutputAdjust( 2308 WhereClause *pWC, /* The WHERE clause */ 2309 WhereLoop *pLoop, /* The loop to adjust downward */ 2310 LogEst nRow /* Number of rows in the entire table */ 2311 ){ 2312 WhereTerm *pTerm, *pX; 2313 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2314 int i, j; 2315 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2316 2317 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2318 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ 2319 assert( pTerm!=0 ); 2320 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; 2321 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2322 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2323 for(j=pLoop->nLTerm-1; j>=0; j--){ 2324 pX = pLoop->aLTerm[j]; 2325 if( pX==0 ) continue; 2326 if( pX==pTerm ) break; 2327 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2328 } 2329 if( j<0 ){ 2330 if( pTerm->truthProb<=0 ){ 2331 /* If a truth probability is specified using the likelihood() hints, 2332 ** then use the probability provided by the application. */ 2333 pLoop->nOut += pTerm->truthProb; 2334 }else{ 2335 /* In the absence of explicit truth probabilities, use heuristics to 2336 ** guess a reasonable truth probability. */ 2337 pLoop->nOut--; 2338 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 2339 && (pTerm->wtFlags & TERM_HIGHTRUTH)==0 /* tag-20200224-1 */ 2340 ){ 2341 Expr *pRight = pTerm->pExpr->pRight; 2342 int k = 0; 2343 testcase( pTerm->pExpr->op==TK_IS ); 2344 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2345 k = 10; 2346 }else{ 2347 k = 20; 2348 } 2349 if( iReduce<k ){ 2350 pTerm->wtFlags |= TERM_HEURTRUTH; 2351 iReduce = k; 2352 } 2353 } 2354 } 2355 } 2356 } 2357 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce; 2358 } 2359 2360 /* 2361 ** Term pTerm is a vector range comparison operation. The first comparison 2362 ** in the vector can be optimized using column nEq of the index. This 2363 ** function returns the total number of vector elements that can be used 2364 ** as part of the range comparison. 2365 ** 2366 ** For example, if the query is: 2367 ** 2368 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2369 ** 2370 ** and the index: 2371 ** 2372 ** CREATE INDEX ... ON (a, b, c, d, e) 2373 ** 2374 ** then this function would be invoked with nEq=1. The value returned in 2375 ** this case is 3. 2376 */ 2377 static int whereRangeVectorLen( 2378 Parse *pParse, /* Parsing context */ 2379 int iCur, /* Cursor open on pIdx */ 2380 Index *pIdx, /* The index to be used for a inequality constraint */ 2381 int nEq, /* Number of prior equality constraints on same index */ 2382 WhereTerm *pTerm /* The vector inequality constraint */ 2383 ){ 2384 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2385 int i; 2386 2387 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2388 for(i=1; i<nCmp; i++){ 2389 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2390 ** of the index. If not, exit the loop. */ 2391 char aff; /* Comparison affinity */ 2392 char idxaff = 0; /* Indexed columns affinity */ 2393 CollSeq *pColl; /* Comparison collation sequence */ 2394 Expr *pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2395 Expr *pRhs = pTerm->pExpr->pRight; 2396 if( pRhs->flags & EP_xIsSelect ){ 2397 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2398 }else{ 2399 pRhs = pRhs->x.pList->a[i].pExpr; 2400 } 2401 2402 /* Check that the LHS of the comparison is a column reference to 2403 ** the right column of the right source table. And that the sort 2404 ** order of the index column is the same as the sort order of the 2405 ** leftmost index column. */ 2406 if( pLhs->op!=TK_COLUMN 2407 || pLhs->iTable!=iCur 2408 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2409 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2410 ){ 2411 break; 2412 } 2413 2414 testcase( pLhs->iColumn==XN_ROWID ); 2415 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2416 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2417 if( aff!=idxaff ) break; 2418 2419 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2420 if( pColl==0 ) break; 2421 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2422 } 2423 return i; 2424 } 2425 2426 /* 2427 ** Adjust the cost C by the costMult facter T. This only occurs if 2428 ** compiled with -DSQLITE_ENABLE_COSTMULT 2429 */ 2430 #ifdef SQLITE_ENABLE_COSTMULT 2431 # define ApplyCostMultiplier(C,T) C += T 2432 #else 2433 # define ApplyCostMultiplier(C,T) 2434 #endif 2435 2436 /* 2437 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2438 ** index pIndex. Try to match one more. 2439 ** 2440 ** When this function is called, pBuilder->pNew->nOut contains the 2441 ** number of rows expected to be visited by filtering using the nEq 2442 ** terms only. If it is modified, this value is restored before this 2443 ** function returns. 2444 ** 2445 ** If pProbe->idxType==SQLITE_IDXTYPE_IPK, that means pIndex is 2446 ** a fake index used for the INTEGER PRIMARY KEY. 2447 */ 2448 static int whereLoopAddBtreeIndex( 2449 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2450 SrcItem *pSrc, /* FROM clause term being analyzed */ 2451 Index *pProbe, /* An index on pSrc */ 2452 LogEst nInMul /* log(Number of iterations due to IN) */ 2453 ){ 2454 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2455 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2456 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2457 WhereLoop *pNew; /* Template WhereLoop under construction */ 2458 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2459 int opMask; /* Valid operators for constraints */ 2460 WhereScan scan; /* Iterator for WHERE terms */ 2461 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2462 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2463 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2464 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2465 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2466 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2467 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2468 LogEst saved_nOut; /* Original value of pNew->nOut */ 2469 int rc = SQLITE_OK; /* Return code */ 2470 LogEst rSize; /* Number of rows in the table */ 2471 LogEst rLogSize; /* Logarithm of table size */ 2472 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2473 2474 pNew = pBuilder->pNew; 2475 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2476 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d, nSkip=%d, rRun=%d\n", 2477 pProbe->pTable->zName,pProbe->zName, 2478 pNew->u.btree.nEq, pNew->nSkip, pNew->rRun)); 2479 2480 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2481 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2482 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2483 opMask = WO_LT|WO_LE; 2484 }else{ 2485 assert( pNew->u.btree.nBtm==0 ); 2486 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2487 } 2488 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2489 2490 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2491 2492 saved_nEq = pNew->u.btree.nEq; 2493 saved_nBtm = pNew->u.btree.nBtm; 2494 saved_nTop = pNew->u.btree.nTop; 2495 saved_nSkip = pNew->nSkip; 2496 saved_nLTerm = pNew->nLTerm; 2497 saved_wsFlags = pNew->wsFlags; 2498 saved_prereq = pNew->prereq; 2499 saved_nOut = pNew->nOut; 2500 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2501 opMask, pProbe); 2502 pNew->rSetup = 0; 2503 rSize = pProbe->aiRowLogEst[0]; 2504 rLogSize = estLog(rSize); 2505 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2506 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2507 LogEst rCostIdx; 2508 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2509 int nIn = 0; 2510 #ifdef SQLITE_ENABLE_STAT4 2511 int nRecValid = pBuilder->nRecValid; 2512 #endif 2513 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2514 && indexColumnNotNull(pProbe, saved_nEq) 2515 ){ 2516 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2517 } 2518 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2519 2520 /* Do not allow the upper bound of a LIKE optimization range constraint 2521 ** to mix with a lower range bound from some other source */ 2522 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2523 2524 /* tag-20191211-001: Do not allow constraints from the WHERE clause to 2525 ** be used by the right table of a LEFT JOIN. Only constraints in the 2526 ** ON clause are allowed. See tag-20191211-002 for the vtab equivalent. */ 2527 if( (pSrc->fg.jointype & JT_LEFT)!=0 2528 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2529 ){ 2530 continue; 2531 } 2532 2533 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2534 pBuilder->bldFlags1 |= SQLITE_BLDF1_UNIQUE; 2535 }else{ 2536 pBuilder->bldFlags1 |= SQLITE_BLDF1_INDEXED; 2537 } 2538 pNew->wsFlags = saved_wsFlags; 2539 pNew->u.btree.nEq = saved_nEq; 2540 pNew->u.btree.nBtm = saved_nBtm; 2541 pNew->u.btree.nTop = saved_nTop; 2542 pNew->nLTerm = saved_nLTerm; 2543 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2544 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2545 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2546 2547 assert( nInMul==0 2548 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2549 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2550 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2551 ); 2552 2553 if( eOp & WO_IN ){ 2554 Expr *pExpr = pTerm->pExpr; 2555 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 2556 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2557 int i; 2558 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2559 2560 /* The expression may actually be of the form (x, y) IN (SELECT...). 2561 ** In this case there is a separate term for each of (x) and (y). 2562 ** However, the nIn multiplier should only be applied once, not once 2563 ** for each such term. The following loop checks that pTerm is the 2564 ** first such term in use, and sets nIn back to 0 if it is not. */ 2565 for(i=0; i<pNew->nLTerm-1; i++){ 2566 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2567 } 2568 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2569 /* "x IN (value, value, ...)" */ 2570 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2571 } 2572 if( pProbe->hasStat1 && rLogSize>=10 ){ 2573 LogEst M, logK, safetyMargin; 2574 /* Let: 2575 ** N = the total number of rows in the table 2576 ** K = the number of entries on the RHS of the IN operator 2577 ** M = the number of rows in the table that match terms to the 2578 ** to the left in the same index. If the IN operator is on 2579 ** the left-most index column, M==N. 2580 ** 2581 ** Given the definitions above, it is better to omit the IN operator 2582 ** from the index lookup and instead do a scan of the M elements, 2583 ** testing each scanned row against the IN operator separately, if: 2584 ** 2585 ** M*log(K) < K*log(N) 2586 ** 2587 ** Our estimates for M, K, and N might be inaccurate, so we build in 2588 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator 2589 ** with the index, as using an index has better worst-case behavior. 2590 ** If we do not have real sqlite_stat1 data, always prefer to use 2591 ** the index. Do not bother with this optimization on very small 2592 ** tables (less than 2 rows) as it is pointless in that case. 2593 */ 2594 M = pProbe->aiRowLogEst[saved_nEq]; 2595 logK = estLog(nIn); 2596 safetyMargin = 10; /* TUNING: extra weight for indexed IN */ 2597 if( M + logK + safetyMargin < nIn + rLogSize ){ 2598 WHERETRACE(0x40, 2599 ("Scan preferred over IN operator on column %d of \"%s\" (%d<%d)\n", 2600 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2601 pNew->wsFlags |= WHERE_IN_SEEKSCAN; 2602 }else{ 2603 WHERETRACE(0x40, 2604 ("IN operator preferred on column %d of \"%s\" (%d>=%d)\n", 2605 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2606 } 2607 } 2608 pNew->wsFlags |= WHERE_COLUMN_IN; 2609 }else if( eOp & (WO_EQ|WO_IS) ){ 2610 int iCol = pProbe->aiColumn[saved_nEq]; 2611 pNew->wsFlags |= WHERE_COLUMN_EQ; 2612 assert( saved_nEq==pNew->u.btree.nEq ); 2613 if( iCol==XN_ROWID 2614 || (iCol>=0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2615 ){ 2616 if( iCol==XN_ROWID || pProbe->uniqNotNull 2617 || (pProbe->nKeyCol==1 && pProbe->onError && eOp==WO_EQ) 2618 ){ 2619 pNew->wsFlags |= WHERE_ONEROW; 2620 }else{ 2621 pNew->wsFlags |= WHERE_UNQ_WANTED; 2622 } 2623 } 2624 }else if( eOp & WO_ISNULL ){ 2625 pNew->wsFlags |= WHERE_COLUMN_NULL; 2626 }else if( eOp & (WO_GT|WO_GE) ){ 2627 testcase( eOp & WO_GT ); 2628 testcase( eOp & WO_GE ); 2629 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2630 pNew->u.btree.nBtm = whereRangeVectorLen( 2631 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2632 ); 2633 pBtm = pTerm; 2634 pTop = 0; 2635 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2636 /* Range constraints that come from the LIKE optimization are 2637 ** always used in pairs. */ 2638 pTop = &pTerm[1]; 2639 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2640 assert( pTop->wtFlags & TERM_LIKEOPT ); 2641 assert( pTop->eOperator==WO_LT ); 2642 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2643 pNew->aLTerm[pNew->nLTerm++] = pTop; 2644 pNew->wsFlags |= WHERE_TOP_LIMIT; 2645 pNew->u.btree.nTop = 1; 2646 } 2647 }else{ 2648 assert( eOp & (WO_LT|WO_LE) ); 2649 testcase( eOp & WO_LT ); 2650 testcase( eOp & WO_LE ); 2651 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2652 pNew->u.btree.nTop = whereRangeVectorLen( 2653 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2654 ); 2655 pTop = pTerm; 2656 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2657 pNew->aLTerm[pNew->nLTerm-2] : 0; 2658 } 2659 2660 /* At this point pNew->nOut is set to the number of rows expected to 2661 ** be visited by the index scan before considering term pTerm, or the 2662 ** values of nIn and nInMul. In other words, assuming that all 2663 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2664 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2665 assert( pNew->nOut==saved_nOut ); 2666 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2667 /* Adjust nOut using stat4 data. Or, if there is no stat4 2668 ** data, using some other estimate. */ 2669 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2670 }else{ 2671 int nEq = ++pNew->u.btree.nEq; 2672 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2673 2674 assert( pNew->nOut==saved_nOut ); 2675 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2676 assert( (eOp & WO_IN) || nIn==0 ); 2677 testcase( eOp & WO_IN ); 2678 pNew->nOut += pTerm->truthProb; 2679 pNew->nOut -= nIn; 2680 }else{ 2681 #ifdef SQLITE_ENABLE_STAT4 2682 tRowcnt nOut = 0; 2683 if( nInMul==0 2684 && pProbe->nSample 2685 && pNew->u.btree.nEq<=pProbe->nSampleCol 2686 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) 2687 && OptimizationEnabled(db, SQLITE_Stat4) 2688 ){ 2689 Expr *pExpr = pTerm->pExpr; 2690 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2691 testcase( eOp & WO_EQ ); 2692 testcase( eOp & WO_IS ); 2693 testcase( eOp & WO_ISNULL ); 2694 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2695 }else{ 2696 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2697 } 2698 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2699 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2700 if( nOut ){ 2701 pNew->nOut = sqlite3LogEst(nOut); 2702 if( nEq==1 2703 /* TUNING: Mark terms as "low selectivity" if they seem likely 2704 ** to be true for half or more of the rows in the table. 2705 ** See tag-202002240-1 */ 2706 && pNew->nOut+10 > pProbe->aiRowLogEst[0] 2707 ){ 2708 #if WHERETRACE_ENABLED /* 0x01 */ 2709 if( sqlite3WhereTrace & 0x01 ){ 2710 sqlite3DebugPrintf( 2711 "STAT4 determines term has low selectivity:\n"); 2712 sqlite3WhereTermPrint(pTerm, 999); 2713 } 2714 #endif 2715 pTerm->wtFlags |= TERM_HIGHTRUTH; 2716 if( pTerm->wtFlags & TERM_HEURTRUTH ){ 2717 /* If the term has previously been used with an assumption of 2718 ** higher selectivity, then set the flag to rerun the 2719 ** loop computations. */ 2720 pBuilder->bldFlags2 |= SQLITE_BLDF2_2NDPASS; 2721 } 2722 } 2723 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2724 pNew->nOut -= nIn; 2725 } 2726 } 2727 if( nOut==0 ) 2728 #endif 2729 { 2730 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2731 if( eOp & WO_ISNULL ){ 2732 /* TUNING: If there is no likelihood() value, assume that a 2733 ** "col IS NULL" expression matches twice as many rows 2734 ** as (col=?). */ 2735 pNew->nOut += 10; 2736 } 2737 } 2738 } 2739 } 2740 2741 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2742 ** it to pNew->rRun, which is currently set to the cost of the index 2743 ** seek only. Then, if this is a non-covering index, add the cost of 2744 ** visiting the rows in the main table. */ 2745 assert( pSrc->pTab->szTabRow>0 ); 2746 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2747 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2748 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2749 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2750 } 2751 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2752 2753 nOutUnadjusted = pNew->nOut; 2754 pNew->rRun += nInMul + nIn; 2755 pNew->nOut += nInMul + nIn; 2756 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 2757 rc = whereLoopInsert(pBuilder, pNew); 2758 2759 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2760 pNew->nOut = saved_nOut; 2761 }else{ 2762 pNew->nOut = nOutUnadjusted; 2763 } 2764 2765 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 2766 && pNew->u.btree.nEq<pProbe->nColumn 2767 ){ 2768 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 2769 } 2770 pNew->nOut = saved_nOut; 2771 #ifdef SQLITE_ENABLE_STAT4 2772 pBuilder->nRecValid = nRecValid; 2773 #endif 2774 } 2775 pNew->prereq = saved_prereq; 2776 pNew->u.btree.nEq = saved_nEq; 2777 pNew->u.btree.nBtm = saved_nBtm; 2778 pNew->u.btree.nTop = saved_nTop; 2779 pNew->nSkip = saved_nSkip; 2780 pNew->wsFlags = saved_wsFlags; 2781 pNew->nOut = saved_nOut; 2782 pNew->nLTerm = saved_nLTerm; 2783 2784 /* Consider using a skip-scan if there are no WHERE clause constraints 2785 ** available for the left-most terms of the index, and if the average 2786 ** number of repeats in the left-most terms is at least 18. 2787 ** 2788 ** The magic number 18 is selected on the basis that scanning 17 rows 2789 ** is almost always quicker than an index seek (even though if the index 2790 ** contains fewer than 2^17 rows we assume otherwise in other parts of 2791 ** the code). And, even if it is not, it should not be too much slower. 2792 ** On the other hand, the extra seeks could end up being significantly 2793 ** more expensive. */ 2794 assert( 42==sqlite3LogEst(18) ); 2795 if( saved_nEq==saved_nSkip 2796 && saved_nEq+1<pProbe->nKeyCol 2797 && saved_nEq==pNew->nLTerm 2798 && pProbe->noSkipScan==0 2799 && pProbe->hasStat1!=0 2800 && OptimizationEnabled(db, SQLITE_SkipScan) 2801 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 2802 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 2803 ){ 2804 LogEst nIter; 2805 pNew->u.btree.nEq++; 2806 pNew->nSkip++; 2807 pNew->aLTerm[pNew->nLTerm++] = 0; 2808 pNew->wsFlags |= WHERE_SKIPSCAN; 2809 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 2810 pNew->nOut -= nIter; 2811 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 2812 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 2813 nIter += 5; 2814 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 2815 pNew->nOut = saved_nOut; 2816 pNew->u.btree.nEq = saved_nEq; 2817 pNew->nSkip = saved_nSkip; 2818 pNew->wsFlags = saved_wsFlags; 2819 } 2820 2821 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n", 2822 pProbe->pTable->zName, pProbe->zName, saved_nEq, rc)); 2823 return rc; 2824 } 2825 2826 /* 2827 ** Return True if it is possible that pIndex might be useful in 2828 ** implementing the ORDER BY clause in pBuilder. 2829 ** 2830 ** Return False if pBuilder does not contain an ORDER BY clause or 2831 ** if there is no way for pIndex to be useful in implementing that 2832 ** ORDER BY clause. 2833 */ 2834 static int indexMightHelpWithOrderBy( 2835 WhereLoopBuilder *pBuilder, 2836 Index *pIndex, 2837 int iCursor 2838 ){ 2839 ExprList *pOB; 2840 ExprList *aColExpr; 2841 int ii, jj; 2842 2843 if( pIndex->bUnordered ) return 0; 2844 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 2845 for(ii=0; ii<pOB->nExpr; ii++){ 2846 Expr *pExpr = sqlite3ExprSkipCollateAndLikely(pOB->a[ii].pExpr); 2847 if( NEVER(pExpr==0) ) continue; 2848 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 2849 if( pExpr->iColumn<0 ) return 1; 2850 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2851 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 2852 } 2853 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 2854 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2855 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 2856 if( sqlite3ExprCompareSkip(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 2857 return 1; 2858 } 2859 } 2860 } 2861 } 2862 return 0; 2863 } 2864 2865 /* Check to see if a partial index with pPartIndexWhere can be used 2866 ** in the current query. Return true if it can be and false if not. 2867 */ 2868 static int whereUsablePartialIndex( 2869 int iTab, /* The table for which we want an index */ 2870 int isLeft, /* True if iTab is the right table of a LEFT JOIN */ 2871 WhereClause *pWC, /* The WHERE clause of the query */ 2872 Expr *pWhere /* The WHERE clause from the partial index */ 2873 ){ 2874 int i; 2875 WhereTerm *pTerm; 2876 Parse *pParse = pWC->pWInfo->pParse; 2877 while( pWhere->op==TK_AND ){ 2878 if( !whereUsablePartialIndex(iTab,isLeft,pWC,pWhere->pLeft) ) return 0; 2879 pWhere = pWhere->pRight; 2880 } 2881 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 2882 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2883 Expr *pExpr; 2884 pExpr = pTerm->pExpr; 2885 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 2886 && (isLeft==0 || ExprHasProperty(pExpr, EP_FromJoin)) 2887 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 2888 ){ 2889 return 1; 2890 } 2891 } 2892 return 0; 2893 } 2894 2895 /* 2896 ** Add all WhereLoop objects for a single table of the join where the table 2897 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 2898 ** a b-tree table, not a virtual table. 2899 ** 2900 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 2901 ** are calculated as follows: 2902 ** 2903 ** For a full scan, assuming the table (or index) contains nRow rows: 2904 ** 2905 ** cost = nRow * 3.0 // full-table scan 2906 ** cost = nRow * K // scan of covering index 2907 ** cost = nRow * (K+3.0) // scan of non-covering index 2908 ** 2909 ** where K is a value between 1.1 and 3.0 set based on the relative 2910 ** estimated average size of the index and table records. 2911 ** 2912 ** For an index scan, where nVisit is the number of index rows visited 2913 ** by the scan, and nSeek is the number of seek operations required on 2914 ** the index b-tree: 2915 ** 2916 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 2917 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 2918 ** 2919 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 2920 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 2921 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 2922 ** 2923 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 2924 ** of uncertainty. For this reason, scoring is designed to pick plans that 2925 ** "do the least harm" if the estimates are inaccurate. For example, a 2926 ** log(nRow) factor is omitted from a non-covering index scan in order to 2927 ** bias the scoring in favor of using an index, since the worst-case 2928 ** performance of using an index is far better than the worst-case performance 2929 ** of a full table scan. 2930 */ 2931 static int whereLoopAddBtree( 2932 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2933 Bitmask mPrereq /* Extra prerequesites for using this table */ 2934 ){ 2935 WhereInfo *pWInfo; /* WHERE analysis context */ 2936 Index *pProbe; /* An index we are evaluating */ 2937 Index sPk; /* A fake index object for the primary key */ 2938 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 2939 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 2940 SrcList *pTabList; /* The FROM clause */ 2941 SrcItem *pSrc; /* The FROM clause btree term to add */ 2942 WhereLoop *pNew; /* Template WhereLoop object */ 2943 int rc = SQLITE_OK; /* Return code */ 2944 int iSortIdx = 1; /* Index number */ 2945 int b; /* A boolean value */ 2946 LogEst rSize; /* number of rows in the table */ 2947 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 2948 WhereClause *pWC; /* The parsed WHERE clause */ 2949 Table *pTab; /* Table being queried */ 2950 2951 pNew = pBuilder->pNew; 2952 pWInfo = pBuilder->pWInfo; 2953 pTabList = pWInfo->pTabList; 2954 pSrc = pTabList->a + pNew->iTab; 2955 pTab = pSrc->pTab; 2956 pWC = pBuilder->pWC; 2957 assert( !IsVirtual(pSrc->pTab) ); 2958 2959 if( pSrc->fg.isIndexedBy ){ 2960 /* An INDEXED BY clause specifies a particular index to use */ 2961 pProbe = pSrc->u2.pIBIndex; 2962 }else if( !HasRowid(pTab) ){ 2963 pProbe = pTab->pIndex; 2964 }else{ 2965 /* There is no INDEXED BY clause. Create a fake Index object in local 2966 ** variable sPk to represent the rowid primary key index. Make this 2967 ** fake index the first in a chain of Index objects with all of the real 2968 ** indices to follow */ 2969 Index *pFirst; /* First of real indices on the table */ 2970 memset(&sPk, 0, sizeof(Index)); 2971 sPk.nKeyCol = 1; 2972 sPk.nColumn = 1; 2973 sPk.aiColumn = &aiColumnPk; 2974 sPk.aiRowLogEst = aiRowEstPk; 2975 sPk.onError = OE_Replace; 2976 sPk.pTable = pTab; 2977 sPk.szIdxRow = pTab->szTabRow; 2978 sPk.idxType = SQLITE_IDXTYPE_IPK; 2979 aiRowEstPk[0] = pTab->nRowLogEst; 2980 aiRowEstPk[1] = 0; 2981 pFirst = pSrc->pTab->pIndex; 2982 if( pSrc->fg.notIndexed==0 ){ 2983 /* The real indices of the table are only considered if the 2984 ** NOT INDEXED qualifier is omitted from the FROM clause */ 2985 sPk.pNext = pFirst; 2986 } 2987 pProbe = &sPk; 2988 } 2989 rSize = pTab->nRowLogEst; 2990 rLogSize = estLog(rSize); 2991 2992 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 2993 /* Automatic indexes */ 2994 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 2995 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 2996 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 2997 && !pSrc->fg.isIndexedBy /* Has no INDEXED BY clause */ 2998 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 2999 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 3000 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 3001 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 3002 ){ 3003 /* Generate auto-index WhereLoops */ 3004 WhereTerm *pTerm; 3005 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 3006 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 3007 if( pTerm->prereqRight & pNew->maskSelf ) continue; 3008 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 3009 pNew->u.btree.nEq = 1; 3010 pNew->nSkip = 0; 3011 pNew->u.btree.pIndex = 0; 3012 pNew->nLTerm = 1; 3013 pNew->aLTerm[0] = pTerm; 3014 /* TUNING: One-time cost for computing the automatic index is 3015 ** estimated to be X*N*log2(N) where N is the number of rows in 3016 ** the table being indexed and where X is 7 (LogEst=28) for normal 3017 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value 3018 ** of X is smaller for views and subqueries so that the query planner 3019 ** will be more aggressive about generating automatic indexes for 3020 ** those objects, since there is no opportunity to add schema 3021 ** indexes on subqueries and views. */ 3022 pNew->rSetup = rLogSize + rSize; 3023 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ 3024 pNew->rSetup += 28; 3025 }else{ 3026 pNew->rSetup -= 10; 3027 } 3028 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 3029 if( pNew->rSetup<0 ) pNew->rSetup = 0; 3030 /* TUNING: Each index lookup yields 20 rows in the table. This 3031 ** is more than the usual guess of 10 rows, since we have no way 3032 ** of knowing how selective the index will ultimately be. It would 3033 ** not be unreasonable to make this value much larger. */ 3034 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 3035 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 3036 pNew->wsFlags = WHERE_AUTO_INDEX; 3037 pNew->prereq = mPrereq | pTerm->prereqRight; 3038 rc = whereLoopInsert(pBuilder, pNew); 3039 } 3040 } 3041 } 3042 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 3043 3044 /* Loop over all indices. If there was an INDEXED BY clause, then only 3045 ** consider index pProbe. */ 3046 for(; rc==SQLITE_OK && pProbe; 3047 pProbe=(pSrc->fg.isIndexedBy ? 0 : pProbe->pNext), iSortIdx++ 3048 ){ 3049 int isLeft = (pSrc->fg.jointype & JT_OUTER)!=0; 3050 if( pProbe->pPartIdxWhere!=0 3051 && !whereUsablePartialIndex(pSrc->iCursor, isLeft, pWC, 3052 pProbe->pPartIdxWhere) 3053 ){ 3054 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 3055 continue; /* Partial index inappropriate for this query */ 3056 } 3057 if( pProbe->bNoQuery ) continue; 3058 rSize = pProbe->aiRowLogEst[0]; 3059 pNew->u.btree.nEq = 0; 3060 pNew->u.btree.nBtm = 0; 3061 pNew->u.btree.nTop = 0; 3062 pNew->nSkip = 0; 3063 pNew->nLTerm = 0; 3064 pNew->iSortIdx = 0; 3065 pNew->rSetup = 0; 3066 pNew->prereq = mPrereq; 3067 pNew->nOut = rSize; 3068 pNew->u.btree.pIndex = pProbe; 3069 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 3070 3071 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 3072 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 3073 if( pProbe->idxType==SQLITE_IDXTYPE_IPK ){ 3074 /* Integer primary key index */ 3075 pNew->wsFlags = WHERE_IPK; 3076 3077 /* Full table scan */ 3078 pNew->iSortIdx = b ? iSortIdx : 0; 3079 /* TUNING: Cost of full table scan is 3.0*N. The 3.0 factor is an 3080 ** extra cost designed to discourage the use of full table scans, 3081 ** since index lookups have better worst-case performance if our 3082 ** stat guesses are wrong. Reduce the 3.0 penalty slightly 3083 ** (to 2.75) if we have valid STAT4 information for the table. 3084 ** At 2.75, a full table scan is preferred over using an index on 3085 ** a column with just two distinct values where each value has about 3086 ** an equal number of appearances. Without STAT4 data, we still want 3087 ** to use an index in that case, since the constraint might be for 3088 ** the scarcer of the two values, and in that case an index lookup is 3089 ** better. 3090 */ 3091 #ifdef SQLITE_ENABLE_STAT4 3092 pNew->rRun = rSize + 16 - 2*((pTab->tabFlags & TF_HasStat4)!=0); 3093 #else 3094 pNew->rRun = rSize + 16; 3095 #endif 3096 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3097 whereLoopOutputAdjust(pWC, pNew, rSize); 3098 rc = whereLoopInsert(pBuilder, pNew); 3099 pNew->nOut = rSize; 3100 if( rc ) break; 3101 }else{ 3102 Bitmask m; 3103 if( pProbe->isCovering ){ 3104 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 3105 m = 0; 3106 }else{ 3107 m = pSrc->colUsed & pProbe->colNotIdxed; 3108 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 3109 } 3110 3111 /* Full scan via index */ 3112 if( b 3113 || !HasRowid(pTab) 3114 || pProbe->pPartIdxWhere!=0 3115 || pSrc->fg.isIndexedBy 3116 || ( m==0 3117 && pProbe->bUnordered==0 3118 && (pProbe->szIdxRow<pTab->szTabRow) 3119 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 3120 && sqlite3GlobalConfig.bUseCis 3121 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 3122 ) 3123 ){ 3124 pNew->iSortIdx = b ? iSortIdx : 0; 3125 3126 /* The cost of visiting the index rows is N*K, where K is 3127 ** between 1.1 and 3.0, depending on the relative sizes of the 3128 ** index and table rows. */ 3129 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 3130 if( m!=0 ){ 3131 /* If this is a non-covering index scan, add in the cost of 3132 ** doing table lookups. The cost will be 3x the number of 3133 ** lookups. Take into account WHERE clause terms that can be 3134 ** satisfied using just the index, and that do not require a 3135 ** table lookup. */ 3136 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 3137 int ii; 3138 int iCur = pSrc->iCursor; 3139 WhereClause *pWC2 = &pWInfo->sWC; 3140 for(ii=0; ii<pWC2->nTerm; ii++){ 3141 WhereTerm *pTerm = &pWC2->a[ii]; 3142 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 3143 break; 3144 } 3145 /* pTerm can be evaluated using just the index. So reduce 3146 ** the expected number of table lookups accordingly */ 3147 if( pTerm->truthProb<=0 ){ 3148 nLookup += pTerm->truthProb; 3149 }else{ 3150 nLookup--; 3151 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 3152 } 3153 } 3154 3155 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 3156 } 3157 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3158 whereLoopOutputAdjust(pWC, pNew, rSize); 3159 rc = whereLoopInsert(pBuilder, pNew); 3160 pNew->nOut = rSize; 3161 if( rc ) break; 3162 } 3163 } 3164 3165 pBuilder->bldFlags1 = 0; 3166 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 3167 if( pBuilder->bldFlags1==SQLITE_BLDF1_INDEXED ){ 3168 /* If a non-unique index is used, or if a prefix of the key for 3169 ** unique index is used (making the index functionally non-unique) 3170 ** then the sqlite_stat1 data becomes important for scoring the 3171 ** plan */ 3172 pTab->tabFlags |= TF_StatsUsed; 3173 } 3174 #ifdef SQLITE_ENABLE_STAT4 3175 sqlite3Stat4ProbeFree(pBuilder->pRec); 3176 pBuilder->nRecValid = 0; 3177 pBuilder->pRec = 0; 3178 #endif 3179 } 3180 return rc; 3181 } 3182 3183 #ifndef SQLITE_OMIT_VIRTUALTABLE 3184 3185 /* 3186 ** Argument pIdxInfo is already populated with all constraints that may 3187 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 3188 ** function marks a subset of those constraints usable, invokes the 3189 ** xBestIndex method and adds the returned plan to pBuilder. 3190 ** 3191 ** A constraint is marked usable if: 3192 ** 3193 ** * Argument mUsable indicates that its prerequisites are available, and 3194 ** 3195 ** * It is not one of the operators specified in the mExclude mask passed 3196 ** as the fourth argument (which in practice is either WO_IN or 0). 3197 ** 3198 ** Argument mPrereq is a mask of tables that must be scanned before the 3199 ** virtual table in question. These are added to the plans prerequisites 3200 ** before it is added to pBuilder. 3201 ** 3202 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3203 ** uses one or more WO_IN terms, or false otherwise. 3204 */ 3205 static int whereLoopAddVirtualOne( 3206 WhereLoopBuilder *pBuilder, 3207 Bitmask mPrereq, /* Mask of tables that must be used. */ 3208 Bitmask mUsable, /* Mask of usable tables */ 3209 u16 mExclude, /* Exclude terms using these operators */ 3210 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3211 u16 mNoOmit, /* Do not omit these constraints */ 3212 int *pbIn /* OUT: True if plan uses an IN(...) op */ 3213 ){ 3214 WhereClause *pWC = pBuilder->pWC; 3215 struct sqlite3_index_constraint *pIdxCons; 3216 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3217 int i; 3218 int mxTerm; 3219 int rc = SQLITE_OK; 3220 WhereLoop *pNew = pBuilder->pNew; 3221 Parse *pParse = pBuilder->pWInfo->pParse; 3222 SrcItem *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3223 int nConstraint = pIdxInfo->nConstraint; 3224 3225 assert( (mUsable & mPrereq)==mPrereq ); 3226 *pbIn = 0; 3227 pNew->prereq = mPrereq; 3228 3229 /* Set the usable flag on the subset of constraints identified by 3230 ** arguments mUsable and mExclude. */ 3231 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3232 for(i=0; i<nConstraint; i++, pIdxCons++){ 3233 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3234 pIdxCons->usable = 0; 3235 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3236 && (pTerm->eOperator & mExclude)==0 3237 ){ 3238 pIdxCons->usable = 1; 3239 } 3240 } 3241 3242 /* Initialize the output fields of the sqlite3_index_info structure */ 3243 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3244 assert( pIdxInfo->needToFreeIdxStr==0 ); 3245 pIdxInfo->idxStr = 0; 3246 pIdxInfo->idxNum = 0; 3247 pIdxInfo->orderByConsumed = 0; 3248 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3249 pIdxInfo->estimatedRows = 25; 3250 pIdxInfo->idxFlags = 0; 3251 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3252 3253 /* Invoke the virtual table xBestIndex() method */ 3254 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3255 if( rc ){ 3256 if( rc==SQLITE_CONSTRAINT ){ 3257 /* If the xBestIndex method returns SQLITE_CONSTRAINT, that means 3258 ** that the particular combination of parameters provided is unusable. 3259 ** Make no entries in the loop table. 3260 */ 3261 WHERETRACE(0xffff, (" ^^^^--- non-viable plan rejected!\n")); 3262 return SQLITE_OK; 3263 } 3264 return rc; 3265 } 3266 3267 mxTerm = -1; 3268 assert( pNew->nLSlot>=nConstraint ); 3269 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 3270 pNew->u.vtab.omitMask = 0; 3271 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3272 for(i=0; i<nConstraint; i++, pIdxCons++){ 3273 int iTerm; 3274 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3275 WhereTerm *pTerm; 3276 int j = pIdxCons->iTermOffset; 3277 if( iTerm>=nConstraint 3278 || j<0 3279 || j>=pWC->nTerm 3280 || pNew->aLTerm[iTerm]!=0 3281 || pIdxCons->usable==0 3282 ){ 3283 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3284 testcase( pIdxInfo->needToFreeIdxStr ); 3285 return SQLITE_ERROR; 3286 } 3287 testcase( iTerm==nConstraint-1 ); 3288 testcase( j==0 ); 3289 testcase( j==pWC->nTerm-1 ); 3290 pTerm = &pWC->a[j]; 3291 pNew->prereq |= pTerm->prereqRight; 3292 assert( iTerm<pNew->nLSlot ); 3293 pNew->aLTerm[iTerm] = pTerm; 3294 if( iTerm>mxTerm ) mxTerm = iTerm; 3295 testcase( iTerm==15 ); 3296 testcase( iTerm==16 ); 3297 if( pUsage[i].omit ){ 3298 if( i<16 && ((1<<i)&mNoOmit)==0 ){ 3299 testcase( i!=iTerm ); 3300 pNew->u.vtab.omitMask |= 1<<iTerm; 3301 }else{ 3302 testcase( i!=iTerm ); 3303 } 3304 } 3305 if( (pTerm->eOperator & WO_IN)!=0 ){ 3306 /* A virtual table that is constrained by an IN clause may not 3307 ** consume the ORDER BY clause because (1) the order of IN terms 3308 ** is not necessarily related to the order of output terms and 3309 ** (2) Multiple outputs from a single IN value will not merge 3310 ** together. */ 3311 pIdxInfo->orderByConsumed = 0; 3312 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3313 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3314 } 3315 } 3316 } 3317 3318 pNew->nLTerm = mxTerm+1; 3319 for(i=0; i<=mxTerm; i++){ 3320 if( pNew->aLTerm[i]==0 ){ 3321 /* The non-zero argvIdx values must be contiguous. Raise an 3322 ** error if they are not */ 3323 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3324 testcase( pIdxInfo->needToFreeIdxStr ); 3325 return SQLITE_ERROR; 3326 } 3327 } 3328 assert( pNew->nLTerm<=pNew->nLSlot ); 3329 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3330 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3331 pIdxInfo->needToFreeIdxStr = 0; 3332 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3333 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3334 pIdxInfo->nOrderBy : 0); 3335 pNew->rSetup = 0; 3336 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3337 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3338 3339 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3340 ** that the scan will visit at most one row. Clear it otherwise. */ 3341 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3342 pNew->wsFlags |= WHERE_ONEROW; 3343 }else{ 3344 pNew->wsFlags &= ~WHERE_ONEROW; 3345 } 3346 rc = whereLoopInsert(pBuilder, pNew); 3347 if( pNew->u.vtab.needFree ){ 3348 sqlite3_free(pNew->u.vtab.idxStr); 3349 pNew->u.vtab.needFree = 0; 3350 } 3351 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3352 *pbIn, (sqlite3_uint64)mPrereq, 3353 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3354 3355 return rc; 3356 } 3357 3358 /* 3359 ** If this function is invoked from within an xBestIndex() callback, it 3360 ** returns a pointer to a buffer containing the name of the collation 3361 ** sequence associated with element iCons of the sqlite3_index_info.aConstraint 3362 ** array. Or, if iCons is out of range or there is no active xBestIndex 3363 ** call, return NULL. 3364 */ 3365 const char *sqlite3_vtab_collation(sqlite3_index_info *pIdxInfo, int iCons){ 3366 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3367 const char *zRet = 0; 3368 if( iCons>=0 && iCons<pIdxInfo->nConstraint ){ 3369 CollSeq *pC = 0; 3370 int iTerm = pIdxInfo->aConstraint[iCons].iTermOffset; 3371 Expr *pX = pHidden->pWC->a[iTerm].pExpr; 3372 if( pX->pLeft ){ 3373 pC = sqlite3ExprCompareCollSeq(pHidden->pParse, pX); 3374 } 3375 zRet = (pC ? pC->zName : sqlite3StrBINARY); 3376 } 3377 return zRet; 3378 } 3379 3380 /* 3381 ** Add all WhereLoop objects for a table of the join identified by 3382 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3383 ** 3384 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3385 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3386 ** entries that occur before the virtual table in the FROM clause and are 3387 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3388 ** mUnusable mask contains all FROM clause entries that occur after the 3389 ** virtual table and are separated from it by at least one LEFT or 3390 ** CROSS JOIN. 3391 ** 3392 ** For example, if the query were: 3393 ** 3394 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3395 ** 3396 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3397 ** 3398 ** All the tables in mPrereq must be scanned before the current virtual 3399 ** table. So any terms for which all prerequisites are satisfied by 3400 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3401 ** Conversely, all tables in mUnusable must be scanned after the current 3402 ** virtual table, so any terms for which the prerequisites overlap with 3403 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3404 */ 3405 static int whereLoopAddVirtual( 3406 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3407 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3408 Bitmask mUnusable /* Tables that must be scanned after this one */ 3409 ){ 3410 int rc = SQLITE_OK; /* Return code */ 3411 WhereInfo *pWInfo; /* WHERE analysis context */ 3412 Parse *pParse; /* The parsing context */ 3413 WhereClause *pWC; /* The WHERE clause */ 3414 SrcItem *pSrc; /* The FROM clause term to search */ 3415 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3416 int nConstraint; /* Number of constraints in p */ 3417 int bIn; /* True if plan uses IN(...) operator */ 3418 WhereLoop *pNew; 3419 Bitmask mBest; /* Tables used by best possible plan */ 3420 u16 mNoOmit; 3421 3422 assert( (mPrereq & mUnusable)==0 ); 3423 pWInfo = pBuilder->pWInfo; 3424 pParse = pWInfo->pParse; 3425 pWC = pBuilder->pWC; 3426 pNew = pBuilder->pNew; 3427 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3428 assert( IsVirtual(pSrc->pTab) ); 3429 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy, 3430 &mNoOmit); 3431 if( p==0 ) return SQLITE_NOMEM_BKPT; 3432 pNew->rSetup = 0; 3433 pNew->wsFlags = WHERE_VIRTUALTABLE; 3434 pNew->nLTerm = 0; 3435 pNew->u.vtab.needFree = 0; 3436 nConstraint = p->nConstraint; 3437 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3438 sqlite3DbFree(pParse->db, p); 3439 return SQLITE_NOMEM_BKPT; 3440 } 3441 3442 /* First call xBestIndex() with all constraints usable. */ 3443 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc->pTab->zName)); 3444 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3445 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn); 3446 3447 /* If the call to xBestIndex() with all terms enabled produced a plan 3448 ** that does not require any source tables (IOW: a plan with mBest==0) 3449 ** and does not use an IN(...) operator, then there is no point in making 3450 ** any further calls to xBestIndex() since they will all return the same 3451 ** result (if the xBestIndex() implementation is sane). */ 3452 if( rc==SQLITE_OK && ((mBest = (pNew->prereq & ~mPrereq))!=0 || bIn) ){ 3453 int seenZero = 0; /* True if a plan with no prereqs seen */ 3454 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3455 Bitmask mPrev = 0; 3456 Bitmask mBestNoIn = 0; 3457 3458 /* If the plan produced by the earlier call uses an IN(...) term, call 3459 ** xBestIndex again, this time with IN(...) terms disabled. */ 3460 if( bIn ){ 3461 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3462 rc = whereLoopAddVirtualOne( 3463 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn); 3464 assert( bIn==0 ); 3465 mBestNoIn = pNew->prereq & ~mPrereq; 3466 if( mBestNoIn==0 ){ 3467 seenZero = 1; 3468 seenZeroNoIN = 1; 3469 } 3470 } 3471 3472 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3473 ** in the set of terms that apply to the current virtual table. */ 3474 while( rc==SQLITE_OK ){ 3475 int i; 3476 Bitmask mNext = ALLBITS; 3477 assert( mNext>0 ); 3478 for(i=0; i<nConstraint; i++){ 3479 Bitmask mThis = ( 3480 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3481 ); 3482 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3483 } 3484 mPrev = mNext; 3485 if( mNext==ALLBITS ) break; 3486 if( mNext==mBest || mNext==mBestNoIn ) continue; 3487 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3488 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3489 rc = whereLoopAddVirtualOne( 3490 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn); 3491 if( pNew->prereq==mPrereq ){ 3492 seenZero = 1; 3493 if( bIn==0 ) seenZeroNoIN = 1; 3494 } 3495 } 3496 3497 /* If the calls to xBestIndex() in the above loop did not find a plan 3498 ** that requires no source tables at all (i.e. one guaranteed to be 3499 ** usable), make a call here with all source tables disabled */ 3500 if( rc==SQLITE_OK && seenZero==0 ){ 3501 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3502 rc = whereLoopAddVirtualOne( 3503 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn); 3504 if( bIn==0 ) seenZeroNoIN = 1; 3505 } 3506 3507 /* If the calls to xBestIndex() have so far failed to find a plan 3508 ** that requires no source tables at all and does not use an IN(...) 3509 ** operator, make a final call to obtain one here. */ 3510 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3511 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3512 rc = whereLoopAddVirtualOne( 3513 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn); 3514 } 3515 } 3516 3517 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3518 sqlite3DbFreeNN(pParse->db, p); 3519 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc->pTab->zName, rc)); 3520 return rc; 3521 } 3522 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3523 3524 /* 3525 ** Add WhereLoop entries to handle OR terms. This works for either 3526 ** btrees or virtual tables. 3527 */ 3528 static int whereLoopAddOr( 3529 WhereLoopBuilder *pBuilder, 3530 Bitmask mPrereq, 3531 Bitmask mUnusable 3532 ){ 3533 WhereInfo *pWInfo = pBuilder->pWInfo; 3534 WhereClause *pWC; 3535 WhereLoop *pNew; 3536 WhereTerm *pTerm, *pWCEnd; 3537 int rc = SQLITE_OK; 3538 int iCur; 3539 WhereClause tempWC; 3540 WhereLoopBuilder sSubBuild; 3541 WhereOrSet sSum, sCur; 3542 SrcItem *pItem; 3543 3544 pWC = pBuilder->pWC; 3545 pWCEnd = pWC->a + pWC->nTerm; 3546 pNew = pBuilder->pNew; 3547 memset(&sSum, 0, sizeof(sSum)); 3548 pItem = pWInfo->pTabList->a + pNew->iTab; 3549 iCur = pItem->iCursor; 3550 3551 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3552 if( (pTerm->eOperator & WO_OR)!=0 3553 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3554 ){ 3555 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3556 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3557 WhereTerm *pOrTerm; 3558 int once = 1; 3559 int i, j; 3560 3561 sSubBuild = *pBuilder; 3562 sSubBuild.pOrderBy = 0; 3563 sSubBuild.pOrSet = &sCur; 3564 3565 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3566 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3567 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3568 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3569 }else if( pOrTerm->leftCursor==iCur ){ 3570 tempWC.pWInfo = pWC->pWInfo; 3571 tempWC.pOuter = pWC; 3572 tempWC.op = TK_AND; 3573 tempWC.nTerm = 1; 3574 tempWC.a = pOrTerm; 3575 sSubBuild.pWC = &tempWC; 3576 }else{ 3577 continue; 3578 } 3579 sCur.n = 0; 3580 #ifdef WHERETRACE_ENABLED 3581 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3582 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3583 if( sqlite3WhereTrace & 0x400 ){ 3584 sqlite3WhereClausePrint(sSubBuild.pWC); 3585 } 3586 #endif 3587 #ifndef SQLITE_OMIT_VIRTUALTABLE 3588 if( IsVirtual(pItem->pTab) ){ 3589 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 3590 }else 3591 #endif 3592 { 3593 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 3594 } 3595 if( rc==SQLITE_OK ){ 3596 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 3597 } 3598 assert( rc==SQLITE_OK || rc==SQLITE_DONE || sCur.n==0 ); 3599 testcase( rc==SQLITE_DONE ); 3600 if( sCur.n==0 ){ 3601 sSum.n = 0; 3602 break; 3603 }else if( once ){ 3604 whereOrMove(&sSum, &sCur); 3605 once = 0; 3606 }else{ 3607 WhereOrSet sPrev; 3608 whereOrMove(&sPrev, &sSum); 3609 sSum.n = 0; 3610 for(i=0; i<sPrev.n; i++){ 3611 for(j=0; j<sCur.n; j++){ 3612 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3613 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3614 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3615 } 3616 } 3617 } 3618 } 3619 pNew->nLTerm = 1; 3620 pNew->aLTerm[0] = pTerm; 3621 pNew->wsFlags = WHERE_MULTI_OR; 3622 pNew->rSetup = 0; 3623 pNew->iSortIdx = 0; 3624 memset(&pNew->u, 0, sizeof(pNew->u)); 3625 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3626 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3627 ** of all sub-scans required by the OR-scan. However, due to rounding 3628 ** errors, it may be that the cost of the OR-scan is equal to its 3629 ** most expensive sub-scan. Add the smallest possible penalty 3630 ** (equivalent to multiplying the cost by 1.07) to ensure that 3631 ** this does not happen. Otherwise, for WHERE clauses such as the 3632 ** following where there is an index on "y": 3633 ** 3634 ** WHERE likelihood(x=?, 0.99) OR y=? 3635 ** 3636 ** the planner may elect to "OR" together a full-table scan and an 3637 ** index lookup. And other similarly odd results. */ 3638 pNew->rRun = sSum.a[i].rRun + 1; 3639 pNew->nOut = sSum.a[i].nOut; 3640 pNew->prereq = sSum.a[i].prereq; 3641 rc = whereLoopInsert(pBuilder, pNew); 3642 } 3643 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3644 } 3645 } 3646 return rc; 3647 } 3648 3649 /* 3650 ** Add all WhereLoop objects for all tables 3651 */ 3652 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3653 WhereInfo *pWInfo = pBuilder->pWInfo; 3654 Bitmask mPrereq = 0; 3655 Bitmask mPrior = 0; 3656 int iTab; 3657 SrcList *pTabList = pWInfo->pTabList; 3658 SrcItem *pItem; 3659 SrcItem *pEnd = &pTabList->a[pWInfo->nLevel]; 3660 sqlite3 *db = pWInfo->pParse->db; 3661 int rc = SQLITE_OK; 3662 WhereLoop *pNew; 3663 3664 /* Loop over the tables in the join, from left to right */ 3665 pNew = pBuilder->pNew; 3666 whereLoopInit(pNew); 3667 pBuilder->iPlanLimit = SQLITE_QUERY_PLANNER_LIMIT; 3668 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3669 Bitmask mUnusable = 0; 3670 pNew->iTab = iTab; 3671 pBuilder->iPlanLimit += SQLITE_QUERY_PLANNER_LIMIT_INCR; 3672 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3673 if( (pItem->fg.jointype & (JT_LEFT|JT_CROSS))!=0 ){ 3674 /* This condition is true when pItem is the FROM clause term on the 3675 ** right-hand-side of a LEFT or CROSS JOIN. */ 3676 mPrereq = mPrior; 3677 }else{ 3678 mPrereq = 0; 3679 } 3680 #ifndef SQLITE_OMIT_VIRTUALTABLE 3681 if( IsVirtual(pItem->pTab) ){ 3682 SrcItem *p; 3683 for(p=&pItem[1]; p<pEnd; p++){ 3684 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3685 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3686 } 3687 } 3688 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 3689 }else 3690 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3691 { 3692 rc = whereLoopAddBtree(pBuilder, mPrereq); 3693 } 3694 if( rc==SQLITE_OK && pBuilder->pWC->hasOr ){ 3695 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 3696 } 3697 mPrior |= pNew->maskSelf; 3698 if( rc || db->mallocFailed ){ 3699 if( rc==SQLITE_DONE ){ 3700 /* We hit the query planner search limit set by iPlanLimit */ 3701 sqlite3_log(SQLITE_WARNING, "abbreviated query algorithm search"); 3702 rc = SQLITE_OK; 3703 }else{ 3704 break; 3705 } 3706 } 3707 } 3708 3709 whereLoopClear(db, pNew); 3710 return rc; 3711 } 3712 3713 /* 3714 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 3715 ** parameters) to see if it outputs rows in the requested ORDER BY 3716 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3717 ** 3718 ** N>0: N terms of the ORDER BY clause are satisfied 3719 ** N==0: No terms of the ORDER BY clause are satisfied 3720 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3721 ** 3722 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3723 ** strict. With GROUP BY and DISTINCT the only requirement is that 3724 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3725 ** and DISTINCT do not require rows to appear in any particular order as long 3726 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3727 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3728 ** pOrderBy terms must be matched in strict left-to-right order. 3729 */ 3730 static i8 wherePathSatisfiesOrderBy( 3731 WhereInfo *pWInfo, /* The WHERE clause */ 3732 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3733 WherePath *pPath, /* The WherePath to check */ 3734 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 3735 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3736 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3737 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3738 ){ 3739 u8 revSet; /* True if rev is known */ 3740 u8 rev; /* Composite sort order */ 3741 u8 revIdx; /* Index sort order */ 3742 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 3743 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 3744 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 3745 u16 eqOpMask; /* Allowed equality operators */ 3746 u16 nKeyCol; /* Number of key columns in pIndex */ 3747 u16 nColumn; /* Total number of ordered columns in the index */ 3748 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 3749 int iLoop; /* Index of WhereLoop in pPath being processed */ 3750 int i, j; /* Loop counters */ 3751 int iCur; /* Cursor number for current WhereLoop */ 3752 int iColumn; /* A column number within table iCur */ 3753 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 3754 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3755 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 3756 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 3757 Index *pIndex; /* The index associated with pLoop */ 3758 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 3759 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 3760 Bitmask obDone; /* Mask of all ORDER BY terms */ 3761 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 3762 Bitmask ready; /* Mask of inner loops */ 3763 3764 /* 3765 ** We say the WhereLoop is "one-row" if it generates no more than one 3766 ** row of output. A WhereLoop is one-row if all of the following are true: 3767 ** (a) All index columns match with WHERE_COLUMN_EQ. 3768 ** (b) The index is unique 3769 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 3770 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 3771 ** 3772 ** We say the WhereLoop is "order-distinct" if the set of columns from 3773 ** that WhereLoop that are in the ORDER BY clause are different for every 3774 ** row of the WhereLoop. Every one-row WhereLoop is automatically 3775 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 3776 ** is not order-distinct. To be order-distinct is not quite the same as being 3777 ** UNIQUE since a UNIQUE column or index can have multiple rows that 3778 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 3779 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 3780 ** 3781 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 3782 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 3783 ** automatically order-distinct. 3784 */ 3785 3786 assert( pOrderBy!=0 ); 3787 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 3788 3789 nOrderBy = pOrderBy->nExpr; 3790 testcase( nOrderBy==BMS-1 ); 3791 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 3792 isOrderDistinct = 1; 3793 obDone = MASKBIT(nOrderBy)-1; 3794 orderDistinctMask = 0; 3795 ready = 0; 3796 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 3797 if( wctrlFlags & (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MAX|WHERE_ORDERBY_MIN) ){ 3798 eqOpMask |= WO_IN; 3799 } 3800 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 3801 if( iLoop>0 ) ready |= pLoop->maskSelf; 3802 if( iLoop<nLoop ){ 3803 pLoop = pPath->aLoop[iLoop]; 3804 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 3805 }else{ 3806 pLoop = pLast; 3807 } 3808 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 3809 if( pLoop->u.vtab.isOrdered && (wctrlFlags & WHERE_DISTINCTBY)==0 ){ 3810 obSat = obDone; 3811 } 3812 break; 3813 }else if( wctrlFlags & WHERE_DISTINCTBY ){ 3814 pLoop->u.btree.nDistinctCol = 0; 3815 } 3816 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 3817 3818 /* Mark off any ORDER BY term X that is a column in the table of 3819 ** the current loop for which there is term in the WHERE 3820 ** clause of the form X IS NULL or X=? that reference only outer 3821 ** loops. 3822 */ 3823 for(i=0; i<nOrderBy; i++){ 3824 if( MASKBIT(i) & obSat ) continue; 3825 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr); 3826 if( NEVER(pOBExpr==0) ) continue; 3827 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue; 3828 if( pOBExpr->iTable!=iCur ) continue; 3829 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 3830 ~ready, eqOpMask, 0); 3831 if( pTerm==0 ) continue; 3832 if( pTerm->eOperator==WO_IN ){ 3833 /* IN terms are only valid for sorting in the ORDER BY LIMIT 3834 ** optimization, and then only if they are actually used 3835 ** by the query plan */ 3836 assert( wctrlFlags & 3837 (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX) ); 3838 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 3839 if( j>=pLoop->nLTerm ) continue; 3840 } 3841 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 3842 Parse *pParse = pWInfo->pParse; 3843 CollSeq *pColl1 = sqlite3ExprNNCollSeq(pParse, pOrderBy->a[i].pExpr); 3844 CollSeq *pColl2 = sqlite3ExprCompareCollSeq(pParse, pTerm->pExpr); 3845 assert( pColl1 ); 3846 if( pColl2==0 || sqlite3StrICmp(pColl1->zName, pColl2->zName) ){ 3847 continue; 3848 } 3849 testcase( pTerm->pExpr->op==TK_IS ); 3850 } 3851 obSat |= MASKBIT(i); 3852 } 3853 3854 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 3855 if( pLoop->wsFlags & WHERE_IPK ){ 3856 pIndex = 0; 3857 nKeyCol = 0; 3858 nColumn = 1; 3859 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 3860 return 0; 3861 }else{ 3862 nKeyCol = pIndex->nKeyCol; 3863 nColumn = pIndex->nColumn; 3864 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 3865 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 3866 || !HasRowid(pIndex->pTable)); 3867 isOrderDistinct = IsUniqueIndex(pIndex) 3868 && (pLoop->wsFlags & WHERE_SKIPSCAN)==0; 3869 } 3870 3871 /* Loop through all columns of the index and deal with the ones 3872 ** that are not constrained by == or IN. 3873 */ 3874 rev = revSet = 0; 3875 distinctColumns = 0; 3876 for(j=0; j<nColumn; j++){ 3877 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 3878 3879 assert( j>=pLoop->u.btree.nEq 3880 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 3881 ); 3882 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 3883 u16 eOp = pLoop->aLTerm[j]->eOperator; 3884 3885 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 3886 ** doing WHERE_ORDERBY_LIMIT processing). Except, IS and ISNULL 3887 ** terms imply that the index is not UNIQUE NOT NULL in which case 3888 ** the loop need to be marked as not order-distinct because it can 3889 ** have repeated NULL rows. 3890 ** 3891 ** If the current term is a column of an ((?,?) IN (SELECT...)) 3892 ** expression for which the SELECT returns more than one column, 3893 ** check that it is the only column used by this loop. Otherwise, 3894 ** if it is one of two or more, none of the columns can be 3895 ** considered to match an ORDER BY term. 3896 */ 3897 if( (eOp & eqOpMask)!=0 ){ 3898 if( eOp & (WO_ISNULL|WO_IS) ){ 3899 testcase( eOp & WO_ISNULL ); 3900 testcase( eOp & WO_IS ); 3901 testcase( isOrderDistinct ); 3902 isOrderDistinct = 0; 3903 } 3904 continue; 3905 }else if( ALWAYS(eOp & WO_IN) ){ 3906 /* ALWAYS() justification: eOp is an equality operator due to the 3907 ** j<pLoop->u.btree.nEq constraint above. Any equality other 3908 ** than WO_IN is captured by the previous "if". So this one 3909 ** always has to be WO_IN. */ 3910 Expr *pX = pLoop->aLTerm[j]->pExpr; 3911 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 3912 if( pLoop->aLTerm[i]->pExpr==pX ){ 3913 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 3914 bOnce = 0; 3915 break; 3916 } 3917 } 3918 } 3919 } 3920 3921 /* Get the column number in the table (iColumn) and sort order 3922 ** (revIdx) for the j-th column of the index. 3923 */ 3924 if( pIndex ){ 3925 iColumn = pIndex->aiColumn[j]; 3926 revIdx = pIndex->aSortOrder[j] & KEYINFO_ORDER_DESC; 3927 if( iColumn==pIndex->pTable->iPKey ) iColumn = XN_ROWID; 3928 }else{ 3929 iColumn = XN_ROWID; 3930 revIdx = 0; 3931 } 3932 3933 /* An unconstrained column that might be NULL means that this 3934 ** WhereLoop is not well-ordered 3935 */ 3936 if( isOrderDistinct 3937 && iColumn>=0 3938 && j>=pLoop->u.btree.nEq 3939 && pIndex->pTable->aCol[iColumn].notNull==0 3940 ){ 3941 isOrderDistinct = 0; 3942 } 3943 3944 /* Find the ORDER BY term that corresponds to the j-th column 3945 ** of the index and mark that ORDER BY term off 3946 */ 3947 isMatch = 0; 3948 for(i=0; bOnce && i<nOrderBy; i++){ 3949 if( MASKBIT(i) & obSat ) continue; 3950 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr); 3951 testcase( wctrlFlags & WHERE_GROUPBY ); 3952 testcase( wctrlFlags & WHERE_DISTINCTBY ); 3953 if( NEVER(pOBExpr==0) ) continue; 3954 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 3955 if( iColumn>=XN_ROWID ){ 3956 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue; 3957 if( pOBExpr->iTable!=iCur ) continue; 3958 if( pOBExpr->iColumn!=iColumn ) continue; 3959 }else{ 3960 Expr *pIdxExpr = pIndex->aColExpr->a[j].pExpr; 3961 if( sqlite3ExprCompareSkip(pOBExpr, pIdxExpr, iCur) ){ 3962 continue; 3963 } 3964 } 3965 if( iColumn!=XN_ROWID ){ 3966 pColl = sqlite3ExprNNCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3967 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 3968 } 3969 if( wctrlFlags & WHERE_DISTINCTBY ){ 3970 pLoop->u.btree.nDistinctCol = j+1; 3971 } 3972 isMatch = 1; 3973 break; 3974 } 3975 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 3976 /* Make sure the sort order is compatible in an ORDER BY clause. 3977 ** Sort order is irrelevant for a GROUP BY clause. */ 3978 if( revSet ){ 3979 if( (rev ^ revIdx)!=(pOrderBy->a[i].sortFlags&KEYINFO_ORDER_DESC) ){ 3980 isMatch = 0; 3981 } 3982 }else{ 3983 rev = revIdx ^ (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC); 3984 if( rev ) *pRevMask |= MASKBIT(iLoop); 3985 revSet = 1; 3986 } 3987 } 3988 if( isMatch && (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL) ){ 3989 if( j==pLoop->u.btree.nEq ){ 3990 pLoop->wsFlags |= WHERE_BIGNULL_SORT; 3991 }else{ 3992 isMatch = 0; 3993 } 3994 } 3995 if( isMatch ){ 3996 if( iColumn==XN_ROWID ){ 3997 testcase( distinctColumns==0 ); 3998 distinctColumns = 1; 3999 } 4000 obSat |= MASKBIT(i); 4001 }else{ 4002 /* No match found */ 4003 if( j==0 || j<nKeyCol ){ 4004 testcase( isOrderDistinct!=0 ); 4005 isOrderDistinct = 0; 4006 } 4007 break; 4008 } 4009 } /* end Loop over all index columns */ 4010 if( distinctColumns ){ 4011 testcase( isOrderDistinct==0 ); 4012 isOrderDistinct = 1; 4013 } 4014 } /* end-if not one-row */ 4015 4016 /* Mark off any other ORDER BY terms that reference pLoop */ 4017 if( isOrderDistinct ){ 4018 orderDistinctMask |= pLoop->maskSelf; 4019 for(i=0; i<nOrderBy; i++){ 4020 Expr *p; 4021 Bitmask mTerm; 4022 if( MASKBIT(i) & obSat ) continue; 4023 p = pOrderBy->a[i].pExpr; 4024 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 4025 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 4026 if( (mTerm&~orderDistinctMask)==0 ){ 4027 obSat |= MASKBIT(i); 4028 } 4029 } 4030 } 4031 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 4032 if( obSat==obDone ) return (i8)nOrderBy; 4033 if( !isOrderDistinct ){ 4034 for(i=nOrderBy-1; i>0; i--){ 4035 Bitmask m = MASKBIT(i) - 1; 4036 if( (obSat&m)==m ) return i; 4037 } 4038 return 0; 4039 } 4040 return -1; 4041 } 4042 4043 4044 /* 4045 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 4046 ** the planner assumes that the specified pOrderBy list is actually a GROUP 4047 ** BY clause - and so any order that groups rows as required satisfies the 4048 ** request. 4049 ** 4050 ** Normally, in this case it is not possible for the caller to determine 4051 ** whether or not the rows are really being delivered in sorted order, or 4052 ** just in some other order that provides the required grouping. However, 4053 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 4054 ** this function may be called on the returned WhereInfo object. It returns 4055 ** true if the rows really will be sorted in the specified order, or false 4056 ** otherwise. 4057 ** 4058 ** For example, assuming: 4059 ** 4060 ** CREATE INDEX i1 ON t1(x, Y); 4061 ** 4062 ** then 4063 ** 4064 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 4065 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 4066 */ 4067 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 4068 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 4069 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 4070 return pWInfo->sorted; 4071 } 4072 4073 #ifdef WHERETRACE_ENABLED 4074 /* For debugging use only: */ 4075 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 4076 static char zName[65]; 4077 int i; 4078 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 4079 if( pLast ) zName[i++] = pLast->cId; 4080 zName[i] = 0; 4081 return zName; 4082 } 4083 #endif 4084 4085 /* 4086 ** Return the cost of sorting nRow rows, assuming that the keys have 4087 ** nOrderby columns and that the first nSorted columns are already in 4088 ** order. 4089 */ 4090 static LogEst whereSortingCost( 4091 WhereInfo *pWInfo, 4092 LogEst nRow, 4093 int nOrderBy, 4094 int nSorted 4095 ){ 4096 /* TUNING: Estimated cost of a full external sort, where N is 4097 ** the number of rows to sort is: 4098 ** 4099 ** cost = (3.0 * N * log(N)). 4100 ** 4101 ** Or, if the order-by clause has X terms but only the last Y 4102 ** terms are out of order, then block-sorting will reduce the 4103 ** sorting cost to: 4104 ** 4105 ** cost = (3.0 * N * log(N)) * (Y/X) 4106 ** 4107 ** The (Y/X) term is implemented using stack variable rScale 4108 ** below. 4109 */ 4110 LogEst rScale, rSortCost; 4111 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 4112 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 4113 rSortCost = nRow + rScale + 16; 4114 4115 /* Multiple by log(M) where M is the number of output rows. 4116 ** Use the LIMIT for M if it is smaller. Or if this sort is for 4117 ** a DISTINCT operator, M will be the number of distinct output 4118 ** rows, so fudge it downwards a bit. 4119 */ 4120 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 4121 nRow = pWInfo->iLimit; 4122 }else if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT) ){ 4123 /* TUNING: In the sort for a DISTINCT operator, assume that the DISTINCT 4124 ** reduces the number of output rows by a factor of 2 */ 4125 if( nRow>10 ) nRow -= 10; assert( 10==sqlite3LogEst(2) ); 4126 } 4127 rSortCost += estLog(nRow); 4128 return rSortCost; 4129 } 4130 4131 /* 4132 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 4133 ** attempts to find the lowest cost path that visits each WhereLoop 4134 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 4135 ** 4136 ** Assume that the total number of output rows that will need to be sorted 4137 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 4138 ** costs if nRowEst==0. 4139 ** 4140 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 4141 ** error occurs. 4142 */ 4143 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 4144 int mxChoice; /* Maximum number of simultaneous paths tracked */ 4145 int nLoop; /* Number of terms in the join */ 4146 Parse *pParse; /* Parsing context */ 4147 sqlite3 *db; /* The database connection */ 4148 int iLoop; /* Loop counter over the terms of the join */ 4149 int ii, jj; /* Loop counters */ 4150 int mxI = 0; /* Index of next entry to replace */ 4151 int nOrderBy; /* Number of ORDER BY clause terms */ 4152 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 4153 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 4154 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 4155 WherePath *aFrom; /* All nFrom paths at the previous level */ 4156 WherePath *aTo; /* The nTo best paths at the current level */ 4157 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 4158 WherePath *pTo; /* An element of aTo[] that we are working on */ 4159 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 4160 WhereLoop **pX; /* Used to divy up the pSpace memory */ 4161 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 4162 char *pSpace; /* Temporary memory used by this routine */ 4163 int nSpace; /* Bytes of space allocated at pSpace */ 4164 4165 pParse = pWInfo->pParse; 4166 db = pParse->db; 4167 nLoop = pWInfo->nLevel; 4168 /* TUNING: For simple queries, only the best path is tracked. 4169 ** For 2-way joins, the 5 best paths are followed. 4170 ** For joins of 3 or more tables, track the 10 best paths */ 4171 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 4172 assert( nLoop<=pWInfo->pTabList->nSrc ); 4173 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 4174 4175 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 4176 ** case the purpose of this call is to estimate the number of rows returned 4177 ** by the overall query. Once this estimate has been obtained, the caller 4178 ** will invoke this function a second time, passing the estimate as the 4179 ** nRowEst parameter. */ 4180 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 4181 nOrderBy = 0; 4182 }else{ 4183 nOrderBy = pWInfo->pOrderBy->nExpr; 4184 } 4185 4186 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 4187 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 4188 nSpace += sizeof(LogEst) * nOrderBy; 4189 pSpace = sqlite3DbMallocRawNN(db, nSpace); 4190 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 4191 aTo = (WherePath*)pSpace; 4192 aFrom = aTo+mxChoice; 4193 memset(aFrom, 0, sizeof(aFrom[0])); 4194 pX = (WhereLoop**)(aFrom+mxChoice); 4195 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 4196 pFrom->aLoop = pX; 4197 } 4198 if( nOrderBy ){ 4199 /* If there is an ORDER BY clause and it is not being ignored, set up 4200 ** space for the aSortCost[] array. Each element of the aSortCost array 4201 ** is either zero - meaning it has not yet been initialized - or the 4202 ** cost of sorting nRowEst rows of data where the first X terms of 4203 ** the ORDER BY clause are already in order, where X is the array 4204 ** index. */ 4205 aSortCost = (LogEst*)pX; 4206 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 4207 } 4208 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 4209 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 4210 4211 /* Seed the search with a single WherePath containing zero WhereLoops. 4212 ** 4213 ** TUNING: Do not let the number of iterations go above 28. If the cost 4214 ** of computing an automatic index is not paid back within the first 28 4215 ** rows, then do not use the automatic index. */ 4216 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 4217 nFrom = 1; 4218 assert( aFrom[0].isOrdered==0 ); 4219 if( nOrderBy ){ 4220 /* If nLoop is zero, then there are no FROM terms in the query. Since 4221 ** in this case the query may return a maximum of one row, the results 4222 ** are already in the requested order. Set isOrdered to nOrderBy to 4223 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 4224 ** -1, indicating that the result set may or may not be ordered, 4225 ** depending on the loops added to the current plan. */ 4226 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 4227 } 4228 4229 /* Compute successively longer WherePaths using the previous generation 4230 ** of WherePaths as the basis for the next. Keep track of the mxChoice 4231 ** best paths at each generation */ 4232 for(iLoop=0; iLoop<nLoop; iLoop++){ 4233 nTo = 0; 4234 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 4235 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 4236 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 4237 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 4238 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 4239 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 4240 Bitmask maskNew; /* Mask of src visited by (..) */ 4241 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 4242 4243 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 4244 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 4245 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<3 ){ 4246 /* Do not use an automatic index if the this loop is expected 4247 ** to run less than 1.25 times. It is tempting to also exclude 4248 ** automatic index usage on an outer loop, but sometimes an automatic 4249 ** index is useful in the outer loop of a correlated subquery. */ 4250 assert( 10==sqlite3LogEst(2) ); 4251 continue; 4252 } 4253 4254 /* At this point, pWLoop is a candidate to be the next loop. 4255 ** Compute its cost */ 4256 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 4257 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 4258 nOut = pFrom->nRow + pWLoop->nOut; 4259 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 4260 if( isOrdered<0 ){ 4261 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 4262 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 4263 iLoop, pWLoop, &revMask); 4264 }else{ 4265 revMask = pFrom->revLoop; 4266 } 4267 if( isOrdered>=0 && isOrdered<nOrderBy ){ 4268 if( aSortCost[isOrdered]==0 ){ 4269 aSortCost[isOrdered] = whereSortingCost( 4270 pWInfo, nRowEst, nOrderBy, isOrdered 4271 ); 4272 } 4273 /* TUNING: Add a small extra penalty (5) to sorting as an 4274 ** extra encouragment to the query planner to select a plan 4275 ** where the rows emerge in the correct order without any sorting 4276 ** required. */ 4277 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]) + 5; 4278 4279 WHERETRACE(0x002, 4280 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 4281 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 4282 rUnsorted, rCost)); 4283 }else{ 4284 rCost = rUnsorted; 4285 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 4286 } 4287 4288 /* Check to see if pWLoop should be added to the set of 4289 ** mxChoice best-so-far paths. 4290 ** 4291 ** First look for an existing path among best-so-far paths 4292 ** that covers the same set of loops and has the same isOrdered 4293 ** setting as the current path candidate. 4294 ** 4295 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 4296 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 4297 ** of legal values for isOrdered, -1..64. 4298 */ 4299 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4300 if( pTo->maskLoop==maskNew 4301 && ((pTo->isOrdered^isOrdered)&0x80)==0 4302 ){ 4303 testcase( jj==nTo-1 ); 4304 break; 4305 } 4306 } 4307 if( jj>=nTo ){ 4308 /* None of the existing best-so-far paths match the candidate. */ 4309 if( nTo>=mxChoice 4310 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4311 ){ 4312 /* The current candidate is no better than any of the mxChoice 4313 ** paths currently in the best-so-far buffer. So discard 4314 ** this candidate as not viable. */ 4315 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4316 if( sqlite3WhereTrace&0x4 ){ 4317 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4318 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4319 isOrdered>=0 ? isOrdered+'0' : '?'); 4320 } 4321 #endif 4322 continue; 4323 } 4324 /* If we reach this points it means that the new candidate path 4325 ** needs to be added to the set of best-so-far paths. */ 4326 if( nTo<mxChoice ){ 4327 /* Increase the size of the aTo set by one */ 4328 jj = nTo++; 4329 }else{ 4330 /* New path replaces the prior worst to keep count below mxChoice */ 4331 jj = mxI; 4332 } 4333 pTo = &aTo[jj]; 4334 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4335 if( sqlite3WhereTrace&0x4 ){ 4336 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4337 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4338 isOrdered>=0 ? isOrdered+'0' : '?'); 4339 } 4340 #endif 4341 }else{ 4342 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4343 ** same set of loops and has the same isOrdered setting as the 4344 ** candidate path. Check to see if the candidate should replace 4345 ** pTo or if the candidate should be skipped. 4346 ** 4347 ** The conditional is an expanded vector comparison equivalent to: 4348 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4349 */ 4350 if( pTo->rCost<rCost 4351 || (pTo->rCost==rCost 4352 && (pTo->nRow<nOut 4353 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4354 ) 4355 ) 4356 ){ 4357 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4358 if( sqlite3WhereTrace&0x4 ){ 4359 sqlite3DebugPrintf( 4360 "Skip %s cost=%-3d,%3d,%3d order=%c", 4361 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4362 isOrdered>=0 ? isOrdered+'0' : '?'); 4363 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4364 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4365 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4366 } 4367 #endif 4368 /* Discard the candidate path from further consideration */ 4369 testcase( pTo->rCost==rCost ); 4370 continue; 4371 } 4372 testcase( pTo->rCost==rCost+1 ); 4373 /* Control reaches here if the candidate path is better than the 4374 ** pTo path. Replace pTo with the candidate. */ 4375 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4376 if( sqlite3WhereTrace&0x4 ){ 4377 sqlite3DebugPrintf( 4378 "Update %s cost=%-3d,%3d,%3d order=%c", 4379 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4380 isOrdered>=0 ? isOrdered+'0' : '?'); 4381 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4382 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4383 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4384 } 4385 #endif 4386 } 4387 /* pWLoop is a winner. Add it to the set of best so far */ 4388 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4389 pTo->revLoop = revMask; 4390 pTo->nRow = nOut; 4391 pTo->rCost = rCost; 4392 pTo->rUnsorted = rUnsorted; 4393 pTo->isOrdered = isOrdered; 4394 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4395 pTo->aLoop[iLoop] = pWLoop; 4396 if( nTo>=mxChoice ){ 4397 mxI = 0; 4398 mxCost = aTo[0].rCost; 4399 mxUnsorted = aTo[0].nRow; 4400 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4401 if( pTo->rCost>mxCost 4402 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4403 ){ 4404 mxCost = pTo->rCost; 4405 mxUnsorted = pTo->rUnsorted; 4406 mxI = jj; 4407 } 4408 } 4409 } 4410 } 4411 } 4412 4413 #ifdef WHERETRACE_ENABLED /* >=2 */ 4414 if( sqlite3WhereTrace & 0x02 ){ 4415 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4416 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4417 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4418 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4419 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4420 if( pTo->isOrdered>0 ){ 4421 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4422 }else{ 4423 sqlite3DebugPrintf("\n"); 4424 } 4425 } 4426 } 4427 #endif 4428 4429 /* Swap the roles of aFrom and aTo for the next generation */ 4430 pFrom = aTo; 4431 aTo = aFrom; 4432 aFrom = pFrom; 4433 nFrom = nTo; 4434 } 4435 4436 if( nFrom==0 ){ 4437 sqlite3ErrorMsg(pParse, "no query solution"); 4438 sqlite3DbFreeNN(db, pSpace); 4439 return SQLITE_ERROR; 4440 } 4441 4442 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4443 pFrom = aFrom; 4444 for(ii=1; ii<nFrom; ii++){ 4445 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4446 } 4447 assert( pWInfo->nLevel==nLoop ); 4448 /* Load the lowest cost path into pWInfo */ 4449 for(iLoop=0; iLoop<nLoop; iLoop++){ 4450 WhereLevel *pLevel = pWInfo->a + iLoop; 4451 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4452 pLevel->iFrom = pWLoop->iTab; 4453 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4454 } 4455 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4456 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4457 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4458 && nRowEst 4459 ){ 4460 Bitmask notUsed; 4461 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4462 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4463 if( rc==pWInfo->pResultSet->nExpr ){ 4464 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4465 } 4466 } 4467 pWInfo->bOrderedInnerLoop = 0; 4468 if( pWInfo->pOrderBy ){ 4469 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4470 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4471 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4472 } 4473 }else{ 4474 pWInfo->nOBSat = pFrom->isOrdered; 4475 pWInfo->revMask = pFrom->revLoop; 4476 if( pWInfo->nOBSat<=0 ){ 4477 pWInfo->nOBSat = 0; 4478 if( nLoop>0 ){ 4479 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4480 if( (wsFlags & WHERE_ONEROW)==0 4481 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4482 ){ 4483 Bitmask m = 0; 4484 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4485 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4486 testcase( wsFlags & WHERE_IPK ); 4487 testcase( wsFlags & WHERE_COLUMN_IN ); 4488 if( rc==pWInfo->pOrderBy->nExpr ){ 4489 pWInfo->bOrderedInnerLoop = 1; 4490 pWInfo->revMask = m; 4491 } 4492 } 4493 } 4494 }else if( nLoop 4495 && pWInfo->nOBSat==1 4496 && (pWInfo->wctrlFlags & (WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX))!=0 4497 ){ 4498 pWInfo->bOrderedInnerLoop = 1; 4499 } 4500 } 4501 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4502 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4503 ){ 4504 Bitmask revMask = 0; 4505 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4506 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4507 ); 4508 assert( pWInfo->sorted==0 ); 4509 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4510 pWInfo->sorted = 1; 4511 pWInfo->revMask = revMask; 4512 } 4513 } 4514 } 4515 4516 4517 pWInfo->nRowOut = pFrom->nRow; 4518 4519 /* Free temporary memory and return success */ 4520 sqlite3DbFreeNN(db, pSpace); 4521 return SQLITE_OK; 4522 } 4523 4524 /* 4525 ** Most queries use only a single table (they are not joins) and have 4526 ** simple == constraints against indexed fields. This routine attempts 4527 ** to plan those simple cases using much less ceremony than the 4528 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4529 ** times for the common case. 4530 ** 4531 ** Return non-zero on success, if this query can be handled by this 4532 ** no-frills query planner. Return zero if this query needs the 4533 ** general-purpose query planner. 4534 */ 4535 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4536 WhereInfo *pWInfo; 4537 SrcItem *pItem; 4538 WhereClause *pWC; 4539 WhereTerm *pTerm; 4540 WhereLoop *pLoop; 4541 int iCur; 4542 int j; 4543 Table *pTab; 4544 Index *pIdx; 4545 4546 pWInfo = pBuilder->pWInfo; 4547 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4548 assert( pWInfo->pTabList->nSrc>=1 ); 4549 pItem = pWInfo->pTabList->a; 4550 pTab = pItem->pTab; 4551 if( IsVirtual(pTab) ) return 0; 4552 if( pItem->fg.isIndexedBy ) return 0; 4553 iCur = pItem->iCursor; 4554 pWC = &pWInfo->sWC; 4555 pLoop = pBuilder->pNew; 4556 pLoop->wsFlags = 0; 4557 pLoop->nSkip = 0; 4558 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0); 4559 if( pTerm ){ 4560 testcase( pTerm->eOperator & WO_IS ); 4561 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4562 pLoop->aLTerm[0] = pTerm; 4563 pLoop->nLTerm = 1; 4564 pLoop->u.btree.nEq = 1; 4565 /* TUNING: Cost of a rowid lookup is 10 */ 4566 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 4567 }else{ 4568 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 4569 int opMask; 4570 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 4571 if( !IsUniqueIndex(pIdx) 4572 || pIdx->pPartIdxWhere!=0 4573 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 4574 ) continue; 4575 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 4576 for(j=0; j<pIdx->nKeyCol; j++){ 4577 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx); 4578 if( pTerm==0 ) break; 4579 testcase( pTerm->eOperator & WO_IS ); 4580 pLoop->aLTerm[j] = pTerm; 4581 } 4582 if( j!=pIdx->nKeyCol ) continue; 4583 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 4584 if( pIdx->isCovering || (pItem->colUsed & pIdx->colNotIdxed)==0 ){ 4585 pLoop->wsFlags |= WHERE_IDX_ONLY; 4586 } 4587 pLoop->nLTerm = j; 4588 pLoop->u.btree.nEq = j; 4589 pLoop->u.btree.pIndex = pIdx; 4590 /* TUNING: Cost of a unique index lookup is 15 */ 4591 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 4592 break; 4593 } 4594 } 4595 if( pLoop->wsFlags ){ 4596 pLoop->nOut = (LogEst)1; 4597 pWInfo->a[0].pWLoop = pLoop; 4598 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 4599 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 4600 pWInfo->a[0].iTabCur = iCur; 4601 pWInfo->nRowOut = 1; 4602 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 4603 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 4604 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4605 } 4606 #ifdef SQLITE_DEBUG 4607 pLoop->cId = '0'; 4608 #endif 4609 return 1; 4610 } 4611 return 0; 4612 } 4613 4614 /* 4615 ** Helper function for exprIsDeterministic(). 4616 */ 4617 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 4618 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 4619 pWalker->eCode = 0; 4620 return WRC_Abort; 4621 } 4622 return WRC_Continue; 4623 } 4624 4625 /* 4626 ** Return true if the expression contains no non-deterministic SQL 4627 ** functions. Do not consider non-deterministic SQL functions that are 4628 ** part of sub-select statements. 4629 */ 4630 static int exprIsDeterministic(Expr *p){ 4631 Walker w; 4632 memset(&w, 0, sizeof(w)); 4633 w.eCode = 1; 4634 w.xExprCallback = exprNodeIsDeterministic; 4635 w.xSelectCallback = sqlite3SelectWalkFail; 4636 sqlite3WalkExpr(&w, p); 4637 return w.eCode; 4638 } 4639 4640 4641 #ifdef WHERETRACE_ENABLED 4642 /* 4643 ** Display all WhereLoops in pWInfo 4644 */ 4645 static void showAllWhereLoops(WhereInfo *pWInfo, WhereClause *pWC){ 4646 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4647 WhereLoop *p; 4648 int i; 4649 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4650 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4651 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4652 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 4653 sqlite3WhereLoopPrint(p, pWC); 4654 } 4655 } 4656 } 4657 # define WHERETRACE_ALL_LOOPS(W,C) showAllWhereLoops(W,C) 4658 #else 4659 # define WHERETRACE_ALL_LOOPS(W,C) 4660 #endif 4661 4662 /* 4663 ** Generate the beginning of the loop used for WHERE clause processing. 4664 ** The return value is a pointer to an opaque structure that contains 4665 ** information needed to terminate the loop. Later, the calling routine 4666 ** should invoke sqlite3WhereEnd() with the return value of this function 4667 ** in order to complete the WHERE clause processing. 4668 ** 4669 ** If an error occurs, this routine returns NULL. 4670 ** 4671 ** The basic idea is to do a nested loop, one loop for each table in 4672 ** the FROM clause of a select. (INSERT and UPDATE statements are the 4673 ** same as a SELECT with only a single table in the FROM clause.) For 4674 ** example, if the SQL is this: 4675 ** 4676 ** SELECT * FROM t1, t2, t3 WHERE ...; 4677 ** 4678 ** Then the code generated is conceptually like the following: 4679 ** 4680 ** foreach row1 in t1 do \ Code generated 4681 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 4682 ** foreach row3 in t3 do / 4683 ** ... 4684 ** end \ Code generated 4685 ** end |-- by sqlite3WhereEnd() 4686 ** end / 4687 ** 4688 ** Note that the loops might not be nested in the order in which they 4689 ** appear in the FROM clause if a different order is better able to make 4690 ** use of indices. Note also that when the IN operator appears in 4691 ** the WHERE clause, it might result in additional nested loops for 4692 ** scanning through all values on the right-hand side of the IN. 4693 ** 4694 ** There are Btree cursors associated with each table. t1 uses cursor 4695 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 4696 ** And so forth. This routine generates code to open those VDBE cursors 4697 ** and sqlite3WhereEnd() generates the code to close them. 4698 ** 4699 ** The code that sqlite3WhereBegin() generates leaves the cursors named 4700 ** in pTabList pointing at their appropriate entries. The [...] code 4701 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 4702 ** data from the various tables of the loop. 4703 ** 4704 ** If the WHERE clause is empty, the foreach loops must each scan their 4705 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 4706 ** the tables have indices and there are terms in the WHERE clause that 4707 ** refer to those indices, a complete table scan can be avoided and the 4708 ** code will run much faster. Most of the work of this routine is checking 4709 ** to see if there are indices that can be used to speed up the loop. 4710 ** 4711 ** Terms of the WHERE clause are also used to limit which rows actually 4712 ** make it to the "..." in the middle of the loop. After each "foreach", 4713 ** terms of the WHERE clause that use only terms in that loop and outer 4714 ** loops are evaluated and if false a jump is made around all subsequent 4715 ** inner loops (or around the "..." if the test occurs within the inner- 4716 ** most loop) 4717 ** 4718 ** OUTER JOINS 4719 ** 4720 ** An outer join of tables t1 and t2 is conceptally coded as follows: 4721 ** 4722 ** foreach row1 in t1 do 4723 ** flag = 0 4724 ** foreach row2 in t2 do 4725 ** start: 4726 ** ... 4727 ** flag = 1 4728 ** end 4729 ** if flag==0 then 4730 ** move the row2 cursor to a null row 4731 ** goto start 4732 ** fi 4733 ** end 4734 ** 4735 ** ORDER BY CLAUSE PROCESSING 4736 ** 4737 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 4738 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 4739 ** if there is one. If there is no ORDER BY clause or if this routine 4740 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 4741 ** 4742 ** The iIdxCur parameter is the cursor number of an index. If 4743 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 4744 ** to use for OR clause processing. The WHERE clause should use this 4745 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 4746 ** the first cursor in an array of cursors for all indices. iIdxCur should 4747 ** be used to compute the appropriate cursor depending on which index is 4748 ** used. 4749 */ 4750 WhereInfo *sqlite3WhereBegin( 4751 Parse *pParse, /* The parser context */ 4752 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 4753 Expr *pWhere, /* The WHERE clause */ 4754 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 4755 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 4756 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 4757 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 4758 ** If WHERE_USE_LIMIT, then the limit amount */ 4759 ){ 4760 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4761 int nTabList; /* Number of elements in pTabList */ 4762 WhereInfo *pWInfo; /* Will become the return value of this function */ 4763 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4764 Bitmask notReady; /* Cursors that are not yet positioned */ 4765 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 4766 WhereMaskSet *pMaskSet; /* The expression mask set */ 4767 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 4768 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 4769 int ii; /* Loop counter */ 4770 sqlite3 *db; /* Database connection */ 4771 int rc; /* Return code */ 4772 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 4773 4774 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 4775 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 4776 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4777 )); 4778 4779 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 4780 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4781 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 4782 4783 /* Variable initialization */ 4784 db = pParse->db; 4785 memset(&sWLB, 0, sizeof(sWLB)); 4786 4787 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 4788 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 4789 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 4790 sWLB.pOrderBy = pOrderBy; 4791 4792 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 4793 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 4794 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 4795 wctrlFlags &= ~WHERE_WANT_DISTINCT; 4796 } 4797 4798 /* The number of tables in the FROM clause is limited by the number of 4799 ** bits in a Bitmask 4800 */ 4801 testcase( pTabList->nSrc==BMS ); 4802 if( pTabList->nSrc>BMS ){ 4803 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 4804 return 0; 4805 } 4806 4807 /* This function normally generates a nested loop for all tables in 4808 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 4809 ** only generate code for the first table in pTabList and assume that 4810 ** any cursors associated with subsequent tables are uninitialized. 4811 */ 4812 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 4813 4814 /* Allocate and initialize the WhereInfo structure that will become the 4815 ** return value. A single allocation is used to store the WhereInfo 4816 ** struct, the contents of WhereInfo.a[], the WhereClause structure 4817 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 4818 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 4819 ** some architectures. Hence the ROUND8() below. 4820 */ 4821 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 4822 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 4823 if( db->mallocFailed ){ 4824 sqlite3DbFree(db, pWInfo); 4825 pWInfo = 0; 4826 goto whereBeginError; 4827 } 4828 pWInfo->pParse = pParse; 4829 pWInfo->pTabList = pTabList; 4830 pWInfo->pOrderBy = pOrderBy; 4831 pWInfo->pWhere = pWhere; 4832 pWInfo->pResultSet = pResultSet; 4833 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 4834 pWInfo->nLevel = nTabList; 4835 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(pParse); 4836 pWInfo->wctrlFlags = wctrlFlags; 4837 pWInfo->iLimit = iAuxArg; 4838 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 4839 memset(&pWInfo->nOBSat, 0, 4840 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 4841 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 4842 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 4843 pMaskSet = &pWInfo->sMaskSet; 4844 sWLB.pWInfo = pWInfo; 4845 sWLB.pWC = &pWInfo->sWC; 4846 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 4847 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 4848 whereLoopInit(sWLB.pNew); 4849 #ifdef SQLITE_DEBUG 4850 sWLB.pNew->cId = '*'; 4851 #endif 4852 4853 /* Split the WHERE clause into separate subexpressions where each 4854 ** subexpression is separated by an AND operator. 4855 */ 4856 initMaskSet(pMaskSet); 4857 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 4858 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 4859 4860 /* Special case: No FROM clause 4861 */ 4862 if( nTabList==0 ){ 4863 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 4864 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4865 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4866 } 4867 ExplainQueryPlan((pParse, 0, "SCAN CONSTANT ROW")); 4868 }else{ 4869 /* Assign a bit from the bitmask to every term in the FROM clause. 4870 ** 4871 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 4872 ** 4873 ** The rule of the previous sentence ensures thta if X is the bitmask for 4874 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 4875 ** Knowing the bitmask for all tables to the left of a left join is 4876 ** important. Ticket #3015. 4877 ** 4878 ** Note that bitmasks are created for all pTabList->nSrc tables in 4879 ** pTabList, not just the first nTabList tables. nTabList is normally 4880 ** equal to pTabList->nSrc but might be shortened to 1 if the 4881 ** WHERE_OR_SUBCLAUSE flag is set. 4882 */ 4883 ii = 0; 4884 do{ 4885 createMask(pMaskSet, pTabList->a[ii].iCursor); 4886 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 4887 }while( (++ii)<pTabList->nSrc ); 4888 #ifdef SQLITE_DEBUG 4889 { 4890 Bitmask mx = 0; 4891 for(ii=0; ii<pTabList->nSrc; ii++){ 4892 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 4893 assert( m>=mx ); 4894 mx = m; 4895 } 4896 } 4897 #endif 4898 } 4899 4900 /* Analyze all of the subexpressions. */ 4901 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 4902 if( db->mallocFailed ) goto whereBeginError; 4903 4904 /* Special case: WHERE terms that do not refer to any tables in the join 4905 ** (constant expressions). Evaluate each such term, and jump over all the 4906 ** generated code if the result is not true. 4907 ** 4908 ** Do not do this if the expression contains non-deterministic functions 4909 ** that are not within a sub-select. This is not strictly required, but 4910 ** preserves SQLite's legacy behaviour in the following two cases: 4911 ** 4912 ** FROM ... WHERE random()>0; -- eval random() once per row 4913 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 4914 */ 4915 for(ii=0; ii<sWLB.pWC->nTerm; ii++){ 4916 WhereTerm *pT = &sWLB.pWC->a[ii]; 4917 if( pT->wtFlags & TERM_VIRTUAL ) continue; 4918 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 4919 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 4920 pT->wtFlags |= TERM_CODED; 4921 } 4922 } 4923 4924 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4925 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 4926 /* The DISTINCT marking is pointless. Ignore it. */ 4927 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4928 }else if( pOrderBy==0 ){ 4929 /* Try to ORDER BY the result set to make distinct processing easier */ 4930 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 4931 pWInfo->pOrderBy = pResultSet; 4932 } 4933 } 4934 4935 /* Construct the WhereLoop objects */ 4936 #if defined(WHERETRACE_ENABLED) 4937 if( sqlite3WhereTrace & 0xffff ){ 4938 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 4939 if( wctrlFlags & WHERE_USE_LIMIT ){ 4940 sqlite3DebugPrintf(", limit: %d", iAuxArg); 4941 } 4942 sqlite3DebugPrintf(")\n"); 4943 if( sqlite3WhereTrace & 0x100 ){ 4944 Select sSelect; 4945 memset(&sSelect, 0, sizeof(sSelect)); 4946 sSelect.selFlags = SF_WhereBegin; 4947 sSelect.pSrc = pTabList; 4948 sSelect.pWhere = pWhere; 4949 sSelect.pOrderBy = pOrderBy; 4950 sSelect.pEList = pResultSet; 4951 sqlite3TreeViewSelect(0, &sSelect, 0); 4952 } 4953 } 4954 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 4955 sqlite3DebugPrintf("---- WHERE clause at start of analysis:\n"); 4956 sqlite3WhereClausePrint(sWLB.pWC); 4957 } 4958 #endif 4959 4960 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 4961 rc = whereLoopAddAll(&sWLB); 4962 if( rc ) goto whereBeginError; 4963 4964 #ifdef SQLITE_ENABLE_STAT4 4965 /* If one or more WhereTerm.truthProb values were used in estimating 4966 ** loop parameters, but then those truthProb values were subsequently 4967 ** changed based on STAT4 information while computing subsequent loops, 4968 ** then we need to rerun the whole loop building process so that all 4969 ** loops will be built using the revised truthProb values. */ 4970 if( sWLB.bldFlags2 & SQLITE_BLDF2_2NDPASS ){ 4971 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC); 4972 WHERETRACE(0xffff, 4973 ("**** Redo all loop computations due to" 4974 " TERM_HIGHTRUTH changes ****\n")); 4975 while( pWInfo->pLoops ){ 4976 WhereLoop *p = pWInfo->pLoops; 4977 pWInfo->pLoops = p->pNextLoop; 4978 whereLoopDelete(db, p); 4979 } 4980 rc = whereLoopAddAll(&sWLB); 4981 if( rc ) goto whereBeginError; 4982 } 4983 #endif 4984 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC); 4985 4986 wherePathSolver(pWInfo, 0); 4987 if( db->mallocFailed ) goto whereBeginError; 4988 if( pWInfo->pOrderBy ){ 4989 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 4990 if( db->mallocFailed ) goto whereBeginError; 4991 } 4992 } 4993 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 4994 pWInfo->revMask = ALLBITS; 4995 } 4996 if( pParse->nErr || db->mallocFailed ){ 4997 goto whereBeginError; 4998 } 4999 #ifdef WHERETRACE_ENABLED 5000 if( sqlite3WhereTrace ){ 5001 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 5002 if( pWInfo->nOBSat>0 ){ 5003 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 5004 } 5005 switch( pWInfo->eDistinct ){ 5006 case WHERE_DISTINCT_UNIQUE: { 5007 sqlite3DebugPrintf(" DISTINCT=unique"); 5008 break; 5009 } 5010 case WHERE_DISTINCT_ORDERED: { 5011 sqlite3DebugPrintf(" DISTINCT=ordered"); 5012 break; 5013 } 5014 case WHERE_DISTINCT_UNORDERED: { 5015 sqlite3DebugPrintf(" DISTINCT=unordered"); 5016 break; 5017 } 5018 } 5019 sqlite3DebugPrintf("\n"); 5020 for(ii=0; ii<pWInfo->nLevel; ii++){ 5021 sqlite3WhereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 5022 } 5023 } 5024 #endif 5025 5026 /* Attempt to omit tables from the join that do not affect the result. 5027 ** For a table to not affect the result, the following must be true: 5028 ** 5029 ** 1) The query must not be an aggregate. 5030 ** 2) The table must be the RHS of a LEFT JOIN. 5031 ** 3) Either the query must be DISTINCT, or else the ON or USING clause 5032 ** must contain a constraint that limits the scan of the table to 5033 ** at most a single row. 5034 ** 4) The table must not be referenced by any part of the query apart 5035 ** from its own USING or ON clause. 5036 ** 5037 ** For example, given: 5038 ** 5039 ** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1); 5040 ** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2); 5041 ** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3); 5042 ** 5043 ** then table t2 can be omitted from the following: 5044 ** 5045 ** SELECT v1, v3 FROM t1 5046 ** LEFT JOIN t2 ON (t1.ipk=t2.ipk) 5047 ** LEFT JOIN t3 ON (t1.ipk=t3.ipk) 5048 ** 5049 ** or from: 5050 ** 5051 ** SELECT DISTINCT v1, v3 FROM t1 5052 ** LEFT JOIN t2 5053 ** LEFT JOIN t3 ON (t1.ipk=t3.ipk) 5054 */ 5055 notReady = ~(Bitmask)0; 5056 if( pWInfo->nLevel>=2 5057 && pResultSet!=0 /* these two combine to guarantee */ 5058 && 0==(wctrlFlags & WHERE_AGG_DISTINCT) /* condition (1) above */ 5059 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 5060 ){ 5061 int i; 5062 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet); 5063 if( sWLB.pOrderBy ){ 5064 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy); 5065 } 5066 for(i=pWInfo->nLevel-1; i>=1; i--){ 5067 WhereTerm *pTerm, *pEnd; 5068 SrcItem *pItem; 5069 pLoop = pWInfo->a[i].pWLoop; 5070 pItem = &pWInfo->pTabList->a[pLoop->iTab]; 5071 if( (pItem->fg.jointype & JT_LEFT)==0 ) continue; 5072 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 5073 && (pLoop->wsFlags & WHERE_ONEROW)==0 5074 ){ 5075 continue; 5076 } 5077 if( (tabUsed & pLoop->maskSelf)!=0 ) continue; 5078 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; 5079 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 5080 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 5081 if( !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 5082 || pTerm->pExpr->iRightJoinTable!=pItem->iCursor 5083 ){ 5084 break; 5085 } 5086 } 5087 } 5088 if( pTerm<pEnd ) continue; 5089 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 5090 notReady &= ~pLoop->maskSelf; 5091 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 5092 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 5093 pTerm->wtFlags |= TERM_CODED; 5094 } 5095 } 5096 if( i!=pWInfo->nLevel-1 ){ 5097 int nByte = (pWInfo->nLevel-1-i) * sizeof(WhereLevel); 5098 memmove(&pWInfo->a[i], &pWInfo->a[i+1], nByte); 5099 } 5100 pWInfo->nLevel--; 5101 nTabList--; 5102 } 5103 } 5104 #if defined(WHERETRACE_ENABLED) 5105 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 5106 sqlite3DebugPrintf("---- WHERE clause at end of analysis:\n"); 5107 sqlite3WhereClausePrint(sWLB.pWC); 5108 } 5109 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 5110 #endif 5111 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 5112 5113 /* If the caller is an UPDATE or DELETE statement that is requesting 5114 ** to use a one-pass algorithm, determine if this is appropriate. 5115 ** 5116 ** A one-pass approach can be used if the caller has requested one 5117 ** and either (a) the scan visits at most one row or (b) each 5118 ** of the following are true: 5119 ** 5120 ** * the caller has indicated that a one-pass approach can be used 5121 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and 5122 ** * the table is not a virtual table, and 5123 ** * either the scan does not use the OR optimization or the caller 5124 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified 5125 ** for DELETE). 5126 ** 5127 ** The last qualification is because an UPDATE statement uses 5128 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can 5129 ** use a one-pass approach, and this is not set accurately for scans 5130 ** that use the OR optimization. 5131 */ 5132 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 5133 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 5134 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 5135 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 5136 assert( !(wsFlags & WHERE_VIRTUALTABLE) || IsVirtual(pTabList->a[0].pTab) ); 5137 if( bOnerow || ( 5138 0!=(wctrlFlags & WHERE_ONEPASS_MULTIROW) 5139 && !IsVirtual(pTabList->a[0].pTab) 5140 && (0==(wsFlags & WHERE_MULTI_OR) || (wctrlFlags & WHERE_DUPLICATES_OK)) 5141 )){ 5142 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 5143 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 5144 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 5145 bFordelete = OPFLAG_FORDELETE; 5146 } 5147 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 5148 } 5149 } 5150 } 5151 5152 /* Open all tables in the pTabList and any indices selected for 5153 ** searching those tables. 5154 */ 5155 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 5156 Table *pTab; /* Table to open */ 5157 int iDb; /* Index of database containing table/index */ 5158 SrcItem *pTabItem; 5159 5160 pTabItem = &pTabList->a[pLevel->iFrom]; 5161 pTab = pTabItem->pTab; 5162 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 5163 pLoop = pLevel->pWLoop; 5164 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 5165 /* Do nothing */ 5166 }else 5167 #ifndef SQLITE_OMIT_VIRTUALTABLE 5168 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 5169 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 5170 int iCur = pTabItem->iCursor; 5171 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 5172 }else if( IsVirtual(pTab) ){ 5173 /* noop */ 5174 }else 5175 #endif 5176 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 5177 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 5178 int op = OP_OpenRead; 5179 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5180 op = OP_OpenWrite; 5181 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 5182 }; 5183 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 5184 assert( pTabItem->iCursor==pLevel->iTabCur ); 5185 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 5186 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 5187 if( pWInfo->eOnePass==ONEPASS_OFF 5188 && pTab->nCol<BMS 5189 && (pTab->tabFlags & (TF_HasGenerated|TF_WithoutRowid))==0 5190 ){ 5191 /* If we know that only a prefix of the record will be used, 5192 ** it is advantageous to reduce the "column count" field in 5193 ** the P4 operand of the OP_OpenRead/Write opcode. */ 5194 Bitmask b = pTabItem->colUsed; 5195 int n = 0; 5196 for(; b; b=b>>1, n++){} 5197 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 5198 assert( n<=pTab->nCol ); 5199 } 5200 #ifdef SQLITE_ENABLE_CURSOR_HINTS 5201 if( pLoop->u.btree.pIndex!=0 ){ 5202 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 5203 }else 5204 #endif 5205 { 5206 sqlite3VdbeChangeP5(v, bFordelete); 5207 } 5208 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5209 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 5210 (const u8*)&pTabItem->colUsed, P4_INT64); 5211 #endif 5212 }else{ 5213 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 5214 } 5215 if( pLoop->wsFlags & WHERE_INDEXED ){ 5216 Index *pIx = pLoop->u.btree.pIndex; 5217 int iIndexCur; 5218 int op = OP_OpenRead; 5219 /* iAuxArg is always set to a positive value if ONEPASS is possible */ 5220 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 5221 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 5222 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 5223 ){ 5224 /* This is one term of an OR-optimization using the PRIMARY KEY of a 5225 ** WITHOUT ROWID table. No need for a separate index */ 5226 iIndexCur = pLevel->iTabCur; 5227 op = 0; 5228 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5229 Index *pJ = pTabItem->pTab->pIndex; 5230 iIndexCur = iAuxArg; 5231 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 5232 while( ALWAYS(pJ) && pJ!=pIx ){ 5233 iIndexCur++; 5234 pJ = pJ->pNext; 5235 } 5236 op = OP_OpenWrite; 5237 pWInfo->aiCurOnePass[1] = iIndexCur; 5238 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 5239 iIndexCur = iAuxArg; 5240 op = OP_ReopenIdx; 5241 }else{ 5242 iIndexCur = pParse->nTab++; 5243 } 5244 pLevel->iIdxCur = iIndexCur; 5245 assert( pIx->pSchema==pTab->pSchema ); 5246 assert( iIndexCur>=0 ); 5247 if( op ){ 5248 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 5249 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 5250 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 5251 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 5252 && (pLoop->wsFlags & WHERE_BIGNULL_SORT)==0 5253 && (pLoop->wsFlags & WHERE_IN_SEEKSCAN)==0 5254 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 5255 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 5256 ){ 5257 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); 5258 } 5259 VdbeComment((v, "%s", pIx->zName)); 5260 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5261 { 5262 u64 colUsed = 0; 5263 int ii, jj; 5264 for(ii=0; ii<pIx->nColumn; ii++){ 5265 jj = pIx->aiColumn[ii]; 5266 if( jj<0 ) continue; 5267 if( jj>63 ) jj = 63; 5268 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 5269 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 5270 } 5271 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 5272 (u8*)&colUsed, P4_INT64); 5273 } 5274 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 5275 } 5276 } 5277 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 5278 } 5279 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 5280 if( db->mallocFailed ) goto whereBeginError; 5281 5282 /* Generate the code to do the search. Each iteration of the for 5283 ** loop below generates code for a single nested loop of the VM 5284 ** program. 5285 */ 5286 for(ii=0; ii<nTabList; ii++){ 5287 int addrExplain; 5288 int wsFlags; 5289 pLevel = &pWInfo->a[ii]; 5290 wsFlags = pLevel->pWLoop->wsFlags; 5291 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 5292 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 5293 constructAutomaticIndex(pParse, &pWInfo->sWC, 5294 &pTabList->a[pLevel->iFrom], notReady, pLevel); 5295 if( db->mallocFailed ) goto whereBeginError; 5296 } 5297 #endif 5298 addrExplain = sqlite3WhereExplainOneScan( 5299 pParse, pTabList, pLevel, wctrlFlags 5300 ); 5301 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 5302 notReady = sqlite3WhereCodeOneLoopStart(pParse,v,pWInfo,ii,pLevel,notReady); 5303 pWInfo->iContinue = pLevel->addrCont; 5304 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 5305 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 5306 } 5307 } 5308 5309 /* Done. */ 5310 VdbeModuleComment((v, "Begin WHERE-core")); 5311 pWInfo->iEndWhere = sqlite3VdbeCurrentAddr(v); 5312 return pWInfo; 5313 5314 /* Jump here if malloc fails */ 5315 whereBeginError: 5316 if( pWInfo ){ 5317 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5318 whereInfoFree(db, pWInfo); 5319 } 5320 return 0; 5321 } 5322 5323 /* 5324 ** Part of sqlite3WhereEnd() will rewrite opcodes to reference the 5325 ** index rather than the main table. In SQLITE_DEBUG mode, we want 5326 ** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine 5327 ** does that. 5328 */ 5329 #ifndef SQLITE_DEBUG 5330 # define OpcodeRewriteTrace(D,K,P) /* no-op */ 5331 #else 5332 # define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P) 5333 static void sqlite3WhereOpcodeRewriteTrace( 5334 sqlite3 *db, 5335 int pc, 5336 VdbeOp *pOp 5337 ){ 5338 if( (db->flags & SQLITE_VdbeAddopTrace)==0 ) return; 5339 sqlite3VdbePrintOp(0, pc, pOp); 5340 } 5341 #endif 5342 5343 /* 5344 ** Generate the end of the WHERE loop. See comments on 5345 ** sqlite3WhereBegin() for additional information. 5346 */ 5347 void sqlite3WhereEnd(WhereInfo *pWInfo){ 5348 Parse *pParse = pWInfo->pParse; 5349 Vdbe *v = pParse->pVdbe; 5350 int i; 5351 WhereLevel *pLevel; 5352 WhereLoop *pLoop; 5353 SrcList *pTabList = pWInfo->pTabList; 5354 sqlite3 *db = pParse->db; 5355 int iEnd = sqlite3VdbeCurrentAddr(v); 5356 5357 /* Generate loop termination code. 5358 */ 5359 VdbeModuleComment((v, "End WHERE-core")); 5360 for(i=pWInfo->nLevel-1; i>=0; i--){ 5361 int addr; 5362 pLevel = &pWInfo->a[i]; 5363 pLoop = pLevel->pWLoop; 5364 if( pLevel->op!=OP_Noop ){ 5365 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5366 int addrSeek = 0; 5367 Index *pIdx; 5368 int n; 5369 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 5370 && i==pWInfo->nLevel-1 /* Ticket [ef9318757b152e3] 2017-10-21 */ 5371 && (pLoop->wsFlags & WHERE_INDEXED)!=0 5372 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 5373 && (n = pLoop->u.btree.nDistinctCol)>0 5374 && pIdx->aiRowLogEst[n]>=36 5375 ){ 5376 int r1 = pParse->nMem+1; 5377 int j, op; 5378 for(j=0; j<n; j++){ 5379 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 5380 } 5381 pParse->nMem += n+1; 5382 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 5383 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 5384 VdbeCoverageIf(v, op==OP_SeekLT); 5385 VdbeCoverageIf(v, op==OP_SeekGT); 5386 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 5387 } 5388 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 5389 /* The common case: Advance to the next row */ 5390 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5391 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 5392 sqlite3VdbeChangeP5(v, pLevel->p5); 5393 VdbeCoverage(v); 5394 VdbeCoverageIf(v, pLevel->op==OP_Next); 5395 VdbeCoverageIf(v, pLevel->op==OP_Prev); 5396 VdbeCoverageIf(v, pLevel->op==OP_VNext); 5397 if( pLevel->regBignull ){ 5398 sqlite3VdbeResolveLabel(v, pLevel->addrBignull); 5399 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, pLevel->regBignull, pLevel->p2-1); 5400 VdbeCoverage(v); 5401 } 5402 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5403 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 5404 #endif 5405 }else{ 5406 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5407 } 5408 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 5409 struct InLoop *pIn; 5410 int j; 5411 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 5412 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 5413 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5414 if( pIn->eEndLoopOp!=OP_Noop ){ 5415 if( pIn->nPrefix ){ 5416 int bEarlyOut = 5417 (pLoop->wsFlags & WHERE_VIRTUALTABLE)==0 5418 && (pLoop->wsFlags & WHERE_IN_EARLYOUT)!=0; 5419 if( pLevel->iLeftJoin ){ 5420 /* For LEFT JOIN queries, cursor pIn->iCur may not have been 5421 ** opened yet. This occurs for WHERE clauses such as 5422 ** "a = ? AND b IN (...)", where the index is on (a, b). If 5423 ** the RHS of the (a=?) is NULL, then the "b IN (...)" may 5424 ** never have been coded, but the body of the loop run to 5425 ** return the null-row. So, if the cursor is not open yet, 5426 ** jump over the OP_Next or OP_Prev instruction about to 5427 ** be coded. */ 5428 sqlite3VdbeAddOp2(v, OP_IfNotOpen, pIn->iCur, 5429 sqlite3VdbeCurrentAddr(v) + 2 + bEarlyOut); 5430 VdbeCoverage(v); 5431 } 5432 if( bEarlyOut ){ 5433 sqlite3VdbeAddOp4Int(v, OP_IfNoHope, pLevel->iIdxCur, 5434 sqlite3VdbeCurrentAddr(v)+2, 5435 pIn->iBase, pIn->nPrefix); 5436 VdbeCoverage(v); 5437 } 5438 } 5439 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 5440 VdbeCoverage(v); 5441 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Prev); 5442 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Next); 5443 } 5444 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 5445 } 5446 } 5447 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 5448 if( pLevel->addrSkip ){ 5449 sqlite3VdbeGoto(v, pLevel->addrSkip); 5450 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 5451 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 5452 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 5453 } 5454 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 5455 if( pLevel->addrLikeRep ){ 5456 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 5457 pLevel->addrLikeRep); 5458 VdbeCoverage(v); 5459 } 5460 #endif 5461 if( pLevel->iLeftJoin ){ 5462 int ws = pLoop->wsFlags; 5463 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 5464 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 5465 if( (ws & WHERE_IDX_ONLY)==0 ){ 5466 assert( pLevel->iTabCur==pTabList->a[pLevel->iFrom].iCursor ); 5467 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iTabCur); 5468 } 5469 if( (ws & WHERE_INDEXED) 5470 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCovidx) 5471 ){ 5472 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 5473 } 5474 if( pLevel->op==OP_Return ){ 5475 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 5476 }else{ 5477 sqlite3VdbeGoto(v, pLevel->addrFirst); 5478 } 5479 sqlite3VdbeJumpHere(v, addr); 5480 } 5481 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 5482 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 5483 } 5484 5485 /* The "break" point is here, just past the end of the outer loop. 5486 ** Set it. 5487 */ 5488 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5489 5490 assert( pWInfo->nLevel<=pTabList->nSrc ); 5491 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5492 int k, last; 5493 VdbeOp *pOp, *pLastOp; 5494 Index *pIdx = 0; 5495 SrcItem *pTabItem = &pTabList->a[pLevel->iFrom]; 5496 Table *pTab = pTabItem->pTab; 5497 assert( pTab!=0 ); 5498 pLoop = pLevel->pWLoop; 5499 5500 /* For a co-routine, change all OP_Column references to the table of 5501 ** the co-routine into OP_Copy of result contained in a register. 5502 ** OP_Rowid becomes OP_Null. 5503 */ 5504 if( pTabItem->fg.viaCoroutine ){ 5505 testcase( pParse->db->mallocFailed ); 5506 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 5507 pTabItem->regResult, 0); 5508 continue; 5509 } 5510 5511 #ifdef SQLITE_ENABLE_EARLY_CURSOR_CLOSE 5512 /* Close all of the cursors that were opened by sqlite3WhereBegin. 5513 ** Except, do not close cursors that will be reused by the OR optimization 5514 ** (WHERE_OR_SUBCLAUSE). And do not close the OP_OpenWrite cursors 5515 ** created for the ONEPASS optimization. 5516 */ 5517 if( (pTab->tabFlags & TF_Ephemeral)==0 5518 && pTab->pSelect==0 5519 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5520 ){ 5521 int ws = pLoop->wsFlags; 5522 if( pWInfo->eOnePass==ONEPASS_OFF && (ws & WHERE_IDX_ONLY)==0 ){ 5523 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); 5524 } 5525 if( (ws & WHERE_INDEXED)!=0 5526 && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0 5527 && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1] 5528 ){ 5529 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); 5530 } 5531 } 5532 #endif 5533 5534 /* If this scan uses an index, make VDBE code substitutions to read data 5535 ** from the index instead of from the table where possible. In some cases 5536 ** this optimization prevents the table from ever being read, which can 5537 ** yield a significant performance boost. 5538 ** 5539 ** Calls to the code generator in between sqlite3WhereBegin and 5540 ** sqlite3WhereEnd will have created code that references the table 5541 ** directly. This loop scans all that code looking for opcodes 5542 ** that reference the table and converts them into opcodes that 5543 ** reference the index. 5544 */ 5545 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 5546 pIdx = pLoop->u.btree.pIndex; 5547 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 5548 pIdx = pLevel->u.pCovidx; 5549 } 5550 if( pIdx 5551 && !db->mallocFailed 5552 ){ 5553 if( pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable) ){ 5554 last = iEnd; 5555 }else{ 5556 last = pWInfo->iEndWhere; 5557 } 5558 k = pLevel->addrBody + 1; 5559 #ifdef SQLITE_DEBUG 5560 if( db->flags & SQLITE_VdbeAddopTrace ){ 5561 printf("TRANSLATE opcodes in range %d..%d\n", k, last-1); 5562 } 5563 /* Proof that the "+1" on the k value above is safe */ 5564 pOp = sqlite3VdbeGetOp(v, k - 1); 5565 assert( pOp->opcode!=OP_Column || pOp->p1!=pLevel->iTabCur ); 5566 assert( pOp->opcode!=OP_Rowid || pOp->p1!=pLevel->iTabCur ); 5567 assert( pOp->opcode!=OP_IfNullRow || pOp->p1!=pLevel->iTabCur ); 5568 #endif 5569 pOp = sqlite3VdbeGetOp(v, k); 5570 pLastOp = pOp + (last - k); 5571 assert( pOp<=pLastOp ); 5572 do{ 5573 if( pOp->p1!=pLevel->iTabCur ){ 5574 /* no-op */ 5575 }else if( pOp->opcode==OP_Column 5576 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC 5577 || pOp->opcode==OP_Offset 5578 #endif 5579 ){ 5580 int x = pOp->p2; 5581 assert( pIdx->pTable==pTab ); 5582 if( !HasRowid(pTab) ){ 5583 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 5584 x = pPk->aiColumn[x]; 5585 assert( x>=0 ); 5586 }else{ 5587 testcase( x!=sqlite3StorageColumnToTable(pTab,x) ); 5588 x = sqlite3StorageColumnToTable(pTab,x); 5589 } 5590 x = sqlite3TableColumnToIndex(pIdx, x); 5591 if( x>=0 ){ 5592 pOp->p2 = x; 5593 pOp->p1 = pLevel->iIdxCur; 5594 OpcodeRewriteTrace(db, k, pOp); 5595 } 5596 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 5597 || pWInfo->eOnePass ); 5598 }else if( pOp->opcode==OP_Rowid ){ 5599 pOp->p1 = pLevel->iIdxCur; 5600 pOp->opcode = OP_IdxRowid; 5601 OpcodeRewriteTrace(db, k, pOp); 5602 }else if( pOp->opcode==OP_IfNullRow ){ 5603 pOp->p1 = pLevel->iIdxCur; 5604 OpcodeRewriteTrace(db, k, pOp); 5605 } 5606 #ifdef SQLITE_DEBUG 5607 k++; 5608 #endif 5609 }while( (++pOp)<pLastOp ); 5610 #ifdef SQLITE_DEBUG 5611 if( db->flags & SQLITE_VdbeAddopTrace ) printf("TRANSLATE complete\n"); 5612 #endif 5613 } 5614 } 5615 5616 /* Undo all Expr node modifications */ 5617 while( pWInfo->pExprMods ){ 5618 WhereExprMod *p = pWInfo->pExprMods; 5619 pWInfo->pExprMods = p->pNext; 5620 memcpy(p->pExpr, &p->orig, sizeof(p->orig)); 5621 sqlite3DbFree(db, p); 5622 } 5623 5624 /* Final cleanup 5625 */ 5626 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5627 whereInfoFree(db, pWInfo); 5628 return; 5629 } 5630