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 assert( pMaskSet->n>0 || pMaskSet->ix[0]<0 ); 238 assert( iCursor>=-1 ); 239 if( pMaskSet->ix[0]==iCursor ){ 240 return 1; 241 } 242 for(i=1; i<pMaskSet->n; i++){ 243 if( pMaskSet->ix[i]==iCursor ){ 244 return MASKBIT(i); 245 } 246 } 247 return 0; 248 } 249 250 /* 251 ** Create a new mask for cursor iCursor. 252 ** 253 ** There is one cursor per table in the FROM clause. The number of 254 ** tables in the FROM clause is limited by a test early in the 255 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] 256 ** array will never overflow. 257 */ 258 static void createMask(WhereMaskSet *pMaskSet, int iCursor){ 259 assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); 260 pMaskSet->ix[pMaskSet->n++] = iCursor; 261 } 262 263 /* 264 ** If the right-hand branch of the expression is a TK_COLUMN, then return 265 ** a pointer to the right-hand branch. Otherwise, return NULL. 266 */ 267 static Expr *whereRightSubexprIsColumn(Expr *p){ 268 p = sqlite3ExprSkipCollateAndLikely(p->pRight); 269 if( ALWAYS(p!=0) && p->op==TK_COLUMN && !ExprHasProperty(p, EP_FixedCol) ){ 270 return p; 271 } 272 return 0; 273 } 274 275 /* 276 ** Advance to the next WhereTerm that matches according to the criteria 277 ** established when the pScan object was initialized by whereScanInit(). 278 ** Return NULL if there are no more matching WhereTerms. 279 */ 280 static WhereTerm *whereScanNext(WhereScan *pScan){ 281 int iCur; /* The cursor on the LHS of the term */ 282 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */ 283 Expr *pX; /* An expression being tested */ 284 WhereClause *pWC; /* Shorthand for pScan->pWC */ 285 WhereTerm *pTerm; /* The term being tested */ 286 int k = pScan->k; /* Where to start scanning */ 287 288 assert( pScan->iEquiv<=pScan->nEquiv ); 289 pWC = pScan->pWC; 290 while(1){ 291 iColumn = pScan->aiColumn[pScan->iEquiv-1]; 292 iCur = pScan->aiCur[pScan->iEquiv-1]; 293 assert( pWC!=0 ); 294 assert( iCur>=0 ); 295 do{ 296 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){ 297 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 || pTerm->leftCursor<0 ); 298 if( pTerm->leftCursor==iCur 299 && pTerm->u.x.leftColumn==iColumn 300 && (iColumn!=XN_EXPR 301 || sqlite3ExprCompareSkip(pTerm->pExpr->pLeft, 302 pScan->pIdxExpr,iCur)==0) 303 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin)) 304 ){ 305 if( (pTerm->eOperator & WO_EQUIV)!=0 306 && pScan->nEquiv<ArraySize(pScan->aiCur) 307 && (pX = whereRightSubexprIsColumn(pTerm->pExpr))!=0 308 ){ 309 int j; 310 for(j=0; j<pScan->nEquiv; j++){ 311 if( pScan->aiCur[j]==pX->iTable 312 && pScan->aiColumn[j]==pX->iColumn ){ 313 break; 314 } 315 } 316 if( j==pScan->nEquiv ){ 317 pScan->aiCur[j] = pX->iTable; 318 pScan->aiColumn[j] = pX->iColumn; 319 pScan->nEquiv++; 320 } 321 } 322 if( (pTerm->eOperator & pScan->opMask)!=0 ){ 323 /* Verify the affinity and collating sequence match */ 324 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){ 325 CollSeq *pColl; 326 Parse *pParse = pWC->pWInfo->pParse; 327 pX = pTerm->pExpr; 328 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){ 329 continue; 330 } 331 assert(pX->pLeft); 332 pColl = sqlite3ExprCompareCollSeq(pParse, pX); 333 if( pColl==0 ) pColl = pParse->db->pDfltColl; 334 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){ 335 continue; 336 } 337 } 338 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0 339 && (pX = pTerm->pExpr->pRight, ALWAYS(pX!=0)) 340 && pX->op==TK_COLUMN 341 && pX->iTable==pScan->aiCur[0] 342 && pX->iColumn==pScan->aiColumn[0] 343 ){ 344 testcase( pTerm->eOperator & WO_IS ); 345 continue; 346 } 347 pScan->pWC = pWC; 348 pScan->k = k+1; 349 #ifdef WHERETRACE_ENABLED 350 if( sqlite3WhereTrace & 0x20000 ){ 351 int ii; 352 sqlite3DebugPrintf("SCAN-TERM %p: nEquiv=%d", 353 pTerm, pScan->nEquiv); 354 for(ii=0; ii<pScan->nEquiv; ii++){ 355 sqlite3DebugPrintf(" {%d:%d}", 356 pScan->aiCur[ii], pScan->aiColumn[ii]); 357 } 358 sqlite3DebugPrintf("\n"); 359 } 360 #endif 361 return pTerm; 362 } 363 } 364 } 365 pWC = pWC->pOuter; 366 k = 0; 367 }while( pWC!=0 ); 368 if( pScan->iEquiv>=pScan->nEquiv ) break; 369 pWC = pScan->pOrigWC; 370 k = 0; 371 pScan->iEquiv++; 372 } 373 return 0; 374 } 375 376 /* 377 ** This is whereScanInit() for the case of an index on an expression. 378 ** It is factored out into a separate tail-recursion subroutine so that 379 ** the normal whereScanInit() routine, which is a high-runner, does not 380 ** need to push registers onto the stack as part of its prologue. 381 */ 382 static SQLITE_NOINLINE WhereTerm *whereScanInitIndexExpr(WhereScan *pScan){ 383 pScan->idxaff = sqlite3ExprAffinity(pScan->pIdxExpr); 384 return whereScanNext(pScan); 385 } 386 387 /* 388 ** Initialize a WHERE clause scanner object. Return a pointer to the 389 ** first match. Return NULL if there are no matches. 390 ** 391 ** The scanner will be searching the WHERE clause pWC. It will look 392 ** for terms of the form "X <op> <expr>" where X is column iColumn of table 393 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx 394 ** must be one of the indexes of table iCur. 395 ** 396 ** The <op> must be one of the operators described by opMask. 397 ** 398 ** If the search is for X and the WHERE clause contains terms of the 399 ** form X=Y then this routine might also return terms of the form 400 ** "Y <op> <expr>". The number of levels of transitivity is limited, 401 ** but is enough to handle most commonly occurring SQL statements. 402 ** 403 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with 404 ** index pIdx. 405 */ 406 static WhereTerm *whereScanInit( 407 WhereScan *pScan, /* The WhereScan object being initialized */ 408 WhereClause *pWC, /* The WHERE clause to be scanned */ 409 int iCur, /* Cursor to scan for */ 410 int iColumn, /* Column to scan for */ 411 u32 opMask, /* Operator(s) to scan for */ 412 Index *pIdx /* Must be compatible with this index */ 413 ){ 414 pScan->pOrigWC = pWC; 415 pScan->pWC = pWC; 416 pScan->pIdxExpr = 0; 417 pScan->idxaff = 0; 418 pScan->zCollName = 0; 419 pScan->opMask = opMask; 420 pScan->k = 0; 421 pScan->aiCur[0] = iCur; 422 pScan->nEquiv = 1; 423 pScan->iEquiv = 1; 424 if( pIdx ){ 425 int j = iColumn; 426 iColumn = pIdx->aiColumn[j]; 427 if( iColumn==pIdx->pTable->iPKey ){ 428 iColumn = XN_ROWID; 429 }else if( iColumn>=0 ){ 430 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; 431 pScan->zCollName = pIdx->azColl[j]; 432 }else if( iColumn==XN_EXPR ){ 433 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr; 434 pScan->zCollName = pIdx->azColl[j]; 435 pScan->aiColumn[0] = XN_EXPR; 436 return whereScanInitIndexExpr(pScan); 437 } 438 }else if( iColumn==XN_EXPR ){ 439 return 0; 440 } 441 pScan->aiColumn[0] = iColumn; 442 return whereScanNext(pScan); 443 } 444 445 /* 446 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 447 ** where X is a reference to the iColumn of table iCur or of index pIdx 448 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by 449 ** the op parameter. Return a pointer to the term. Return 0 if not found. 450 ** 451 ** If pIdx!=0 then it must be one of the indexes of table iCur. 452 ** Search for terms matching the iColumn-th column of pIdx 453 ** rather than the iColumn-th column of table iCur. 454 ** 455 ** The term returned might by Y=<expr> if there is another constraint in 456 ** the WHERE clause that specifies that X=Y. Any such constraints will be 457 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 458 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11 459 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10 460 ** other equivalent values. Hence a search for X will return <expr> if X=A1 461 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>. 462 ** 463 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 464 ** then try for the one with no dependencies on <expr> - in other words where 465 ** <expr> is a constant expression of some kind. Only return entries of 466 ** the form "X <op> Y" where Y is a column in another table if no terms of 467 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 468 ** exist, try to return a term that does not use WO_EQUIV. 469 */ 470 WhereTerm *sqlite3WhereFindTerm( 471 WhereClause *pWC, /* The WHERE clause to be searched */ 472 int iCur, /* Cursor number of LHS */ 473 int iColumn, /* Column number of LHS */ 474 Bitmask notReady, /* RHS must not overlap with this mask */ 475 u32 op, /* Mask of WO_xx values describing operator */ 476 Index *pIdx /* Must be compatible with this index, if not NULL */ 477 ){ 478 WhereTerm *pResult = 0; 479 WhereTerm *p; 480 WhereScan scan; 481 482 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); 483 op &= WO_EQ|WO_IS; 484 while( p ){ 485 if( (p->prereqRight & notReady)==0 ){ 486 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){ 487 testcase( p->eOperator & WO_IS ); 488 return p; 489 } 490 if( pResult==0 ) pResult = p; 491 } 492 p = whereScanNext(&scan); 493 } 494 return pResult; 495 } 496 497 /* 498 ** This function searches pList for an entry that matches the iCol-th column 499 ** of index pIdx. 500 ** 501 ** If such an expression is found, its index in pList->a[] is returned. If 502 ** no expression is found, -1 is returned. 503 */ 504 static int findIndexCol( 505 Parse *pParse, /* Parse context */ 506 ExprList *pList, /* Expression list to search */ 507 int iBase, /* Cursor for table associated with pIdx */ 508 Index *pIdx, /* Index to match column of */ 509 int iCol /* Column of index to match */ 510 ){ 511 int i; 512 const char *zColl = pIdx->azColl[iCol]; 513 514 for(i=0; i<pList->nExpr; i++){ 515 Expr *p = sqlite3ExprSkipCollateAndLikely(pList->a[i].pExpr); 516 if( ALWAYS(p!=0) 517 && (p->op==TK_COLUMN || p->op==TK_AGG_COLUMN) 518 && p->iColumn==pIdx->aiColumn[iCol] 519 && p->iTable==iBase 520 ){ 521 CollSeq *pColl = sqlite3ExprNNCollSeq(pParse, pList->a[i].pExpr); 522 if( 0==sqlite3StrICmp(pColl->zName, zColl) ){ 523 return i; 524 } 525 } 526 } 527 528 return -1; 529 } 530 531 /* 532 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL 533 */ 534 static int indexColumnNotNull(Index *pIdx, int iCol){ 535 int j; 536 assert( pIdx!=0 ); 537 assert( iCol>=0 && iCol<pIdx->nColumn ); 538 j = pIdx->aiColumn[iCol]; 539 if( j>=0 ){ 540 return pIdx->pTable->aCol[j].notNull; 541 }else if( j==(-1) ){ 542 return 1; 543 }else{ 544 assert( j==(-2) ); 545 return 0; /* Assume an indexed expression can always yield a NULL */ 546 547 } 548 } 549 550 /* 551 ** Return true if the DISTINCT expression-list passed as the third argument 552 ** is redundant. 553 ** 554 ** A DISTINCT list is redundant if any subset of the columns in the 555 ** DISTINCT list are collectively unique and individually non-null. 556 */ 557 static int isDistinctRedundant( 558 Parse *pParse, /* Parsing context */ 559 SrcList *pTabList, /* The FROM clause */ 560 WhereClause *pWC, /* The WHERE clause */ 561 ExprList *pDistinct /* The result set that needs to be DISTINCT */ 562 ){ 563 Table *pTab; 564 Index *pIdx; 565 int i; 566 int iBase; 567 568 /* If there is more than one table or sub-select in the FROM clause of 569 ** this query, then it will not be possible to show that the DISTINCT 570 ** clause is redundant. */ 571 if( pTabList->nSrc!=1 ) return 0; 572 iBase = pTabList->a[0].iCursor; 573 pTab = pTabList->a[0].pTab; 574 575 /* If any of the expressions is an IPK column on table iBase, then return 576 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 577 ** current SELECT is a correlated sub-query. 578 */ 579 for(i=0; i<pDistinct->nExpr; i++){ 580 Expr *p = sqlite3ExprSkipCollateAndLikely(pDistinct->a[i].pExpr); 581 if( NEVER(p==0) ) continue; 582 if( p->op!=TK_COLUMN && p->op!=TK_AGG_COLUMN ) continue; 583 if( p->iTable==iBase && p->iColumn<0 ) return 1; 584 } 585 586 /* Loop through all indices on the table, checking each to see if it makes 587 ** the DISTINCT qualifier redundant. It does so if: 588 ** 589 ** 1. The index is itself UNIQUE, and 590 ** 591 ** 2. All of the columns in the index are either part of the pDistinct 592 ** list, or else the WHERE clause contains a term of the form "col=X", 593 ** where X is a constant value. The collation sequences of the 594 ** comparison and select-list expressions must match those of the index. 595 ** 596 ** 3. All of those index columns for which the WHERE clause does not 597 ** contain a "col=X" term are subject to a NOT NULL constraint. 598 */ 599 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 600 if( !IsUniqueIndex(pIdx) ) continue; 601 if( pIdx->pPartIdxWhere ) continue; 602 for(i=0; i<pIdx->nKeyCol; i++){ 603 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){ 604 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break; 605 if( indexColumnNotNull(pIdx, i)==0 ) break; 606 } 607 } 608 if( i==pIdx->nKeyCol ){ 609 /* This index implies that the DISTINCT qualifier is redundant. */ 610 return 1; 611 } 612 } 613 614 return 0; 615 } 616 617 618 /* 619 ** Estimate the logarithm of the input value to base 2. 620 */ 621 static LogEst estLog(LogEst N){ 622 return N<=10 ? 0 : sqlite3LogEst(N) - 33; 623 } 624 625 /* 626 ** Convert OP_Column opcodes to OP_Copy in previously generated code. 627 ** 628 ** This routine runs over generated VDBE code and translates OP_Column 629 ** opcodes into OP_Copy when the table is being accessed via co-routine 630 ** instead of via table lookup. 631 ** 632 ** If the iAutoidxCur is not zero, then any OP_Rowid instructions on 633 ** cursor iTabCur are transformed into OP_Sequence opcode for the 634 ** iAutoidxCur cursor, in order to generate unique rowids for the 635 ** automatic index being generated. 636 */ 637 static void translateColumnToCopy( 638 Parse *pParse, /* Parsing context */ 639 int iStart, /* Translate from this opcode to the end */ 640 int iTabCur, /* OP_Column/OP_Rowid references to this table */ 641 int iRegister, /* The first column is in this register */ 642 int iAutoidxCur /* If non-zero, cursor of autoindex being generated */ 643 ){ 644 Vdbe *v = pParse->pVdbe; 645 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart); 646 int iEnd = sqlite3VdbeCurrentAddr(v); 647 if( pParse->db->mallocFailed ) return; 648 for(; iStart<iEnd; iStart++, pOp++){ 649 if( pOp->p1!=iTabCur ) continue; 650 if( pOp->opcode==OP_Column ){ 651 pOp->opcode = OP_Copy; 652 pOp->p1 = pOp->p2 + iRegister; 653 pOp->p2 = pOp->p3; 654 pOp->p3 = 0; 655 }else if( pOp->opcode==OP_Rowid ){ 656 pOp->opcode = OP_Sequence; 657 pOp->p1 = iAutoidxCur; 658 #ifdef SQLITE_ALLOW_ROWID_IN_VIEW 659 if( iAutoidxCur==0 ){ 660 pOp->opcode = OP_Null; 661 pOp->p3 = 0; 662 } 663 #endif 664 } 665 } 666 } 667 668 /* 669 ** Two routines for printing the content of an sqlite3_index_info 670 ** structure. Used for testing and debugging only. If neither 671 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 672 ** are no-ops. 673 */ 674 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) 675 static void whereTraceIndexInfoInputs(sqlite3_index_info *p){ 676 int i; 677 if( !sqlite3WhereTrace ) return; 678 for(i=0; i<p->nConstraint; i++){ 679 sqlite3DebugPrintf( 680 " constraint[%d]: col=%d termid=%d op=%d usabled=%d collseq=%s\n", 681 i, 682 p->aConstraint[i].iColumn, 683 p->aConstraint[i].iTermOffset, 684 p->aConstraint[i].op, 685 p->aConstraint[i].usable, 686 sqlite3_vtab_collation(p,i)); 687 } 688 for(i=0; i<p->nOrderBy; i++){ 689 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 690 i, 691 p->aOrderBy[i].iColumn, 692 p->aOrderBy[i].desc); 693 } 694 } 695 static void whereTraceIndexInfoOutputs(sqlite3_index_info *p){ 696 int i; 697 if( !sqlite3WhereTrace ) return; 698 for(i=0; i<p->nConstraint; i++){ 699 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 700 i, 701 p->aConstraintUsage[i].argvIndex, 702 p->aConstraintUsage[i].omit); 703 } 704 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 705 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 706 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 707 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 708 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); 709 } 710 #else 711 #define whereTraceIndexInfoInputs(A) 712 #define whereTraceIndexInfoOutputs(A) 713 #endif 714 715 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 716 /* 717 ** Return TRUE if the WHERE clause term pTerm is of a form where it 718 ** could be used with an index to access pSrc, assuming an appropriate 719 ** index existed. 720 */ 721 static int termCanDriveIndex( 722 const WhereTerm *pTerm, /* WHERE clause term to check */ 723 const SrcItem *pSrc, /* Table we are trying to access */ 724 const Bitmask notReady /* Tables in outer loops of the join */ 725 ){ 726 char aff; 727 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 728 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0; 729 if( (pSrc->fg.jointype & JT_LEFT) 730 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 731 && (pTerm->eOperator & WO_IS) 732 ){ 733 /* Cannot use an IS term from the WHERE clause as an index driver for 734 ** the RHS of a LEFT JOIN. Such a term can only be used if it is from 735 ** the ON clause. */ 736 return 0; 737 } 738 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 739 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 740 if( pTerm->u.x.leftColumn<0 ) return 0; 741 aff = pSrc->pTab->aCol[pTerm->u.x.leftColumn].affinity; 742 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 743 testcase( pTerm->pExpr->op==TK_IS ); 744 return 1; 745 } 746 #endif 747 748 749 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 750 /* 751 ** Generate code to construct the Index object for an automatic index 752 ** and to set up the WhereLevel object pLevel so that the code generator 753 ** makes use of the automatic index. 754 */ 755 static SQLITE_NOINLINE void constructAutomaticIndex( 756 Parse *pParse, /* The parsing context */ 757 const WhereClause *pWC, /* The WHERE clause */ 758 const SrcItem *pSrc, /* The FROM clause term to get the next index */ 759 const Bitmask notReady, /* Mask of cursors that are not available */ 760 WhereLevel *pLevel /* Write new index here */ 761 ){ 762 int nKeyCol; /* Number of columns in the constructed index */ 763 WhereTerm *pTerm; /* A single term of the WHERE clause */ 764 WhereTerm *pWCEnd; /* End of pWC->a[] */ 765 Index *pIdx; /* Object describing the transient index */ 766 Vdbe *v; /* Prepared statement under construction */ 767 int addrInit; /* Address of the initialization bypass jump */ 768 Table *pTable; /* The table being indexed */ 769 int addrTop; /* Top of the index fill loop */ 770 int regRecord; /* Register holding an index record */ 771 int n; /* Column counter */ 772 int i; /* Loop counter */ 773 int mxBitCol; /* Maximum column in pSrc->colUsed */ 774 CollSeq *pColl; /* Collating sequence to on a column */ 775 WhereLoop *pLoop; /* The Loop object */ 776 char *zNotUsed; /* Extra space on the end of pIdx */ 777 Bitmask idxCols; /* Bitmap of columns used for indexing */ 778 Bitmask extraCols; /* Bitmap of additional columns */ 779 u8 sentWarning = 0; /* True if a warnning has been issued */ 780 Expr *pPartial = 0; /* Partial Index Expression */ 781 int iContinue = 0; /* Jump here to skip excluded rows */ 782 SrcItem *pTabItem; /* FROM clause term being indexed */ 783 int addrCounter = 0; /* Address where integer counter is initialized */ 784 int regBase; /* Array of registers where record is assembled */ 785 786 /* Generate code to skip over the creation and initialization of the 787 ** transient index on 2nd and subsequent iterations of the loop. */ 788 v = pParse->pVdbe; 789 assert( v!=0 ); 790 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 791 792 /* Count the number of columns that will be added to the index 793 ** and used to match WHERE clause constraints */ 794 nKeyCol = 0; 795 pTable = pSrc->pTab; 796 pWCEnd = &pWC->a[pWC->nTerm]; 797 pLoop = pLevel->pWLoop; 798 idxCols = 0; 799 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 800 Expr *pExpr = pTerm->pExpr; 801 /* Make the automatic index a partial index if there are terms in the 802 ** WHERE clause (or the ON clause of a LEFT join) that constrain which 803 ** rows of the target table (pSrc) that can be used. */ 804 if( (pTerm->wtFlags & TERM_VIRTUAL)==0 805 && ((pSrc->fg.jointype&JT_LEFT)==0 || ExprHasProperty(pExpr,EP_FromJoin)) 806 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) 807 ){ 808 pPartial = sqlite3ExprAnd(pParse, pPartial, 809 sqlite3ExprDup(pParse->db, pExpr, 0)); 810 } 811 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 812 int iCol; 813 Bitmask cMask; 814 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 815 iCol = pTerm->u.x.leftColumn; 816 cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 817 testcase( iCol==BMS ); 818 testcase( iCol==BMS-1 ); 819 if( !sentWarning ){ 820 sqlite3_log(SQLITE_WARNING_AUTOINDEX, 821 "automatic index on %s(%s)", pTable->zName, 822 pTable->aCol[iCol].zCnName); 823 sentWarning = 1; 824 } 825 if( (idxCols & cMask)==0 ){ 826 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ 827 goto end_auto_index_create; 828 } 829 pLoop->aLTerm[nKeyCol++] = pTerm; 830 idxCols |= cMask; 831 } 832 } 833 } 834 assert( nKeyCol>0 || pParse->db->mallocFailed ); 835 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; 836 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED 837 | WHERE_AUTO_INDEX; 838 839 /* Count the number of additional columns needed to create a 840 ** covering index. A "covering index" is an index that contains all 841 ** columns that are needed by the query. With a covering index, the 842 ** original table never needs to be accessed. Automatic indices must 843 ** be a covering index because the index will not be updated if the 844 ** original table changes and the index and table cannot both be used 845 ** if they go out of sync. 846 */ 847 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); 848 mxBitCol = MIN(BMS-1,pTable->nCol); 849 testcase( pTable->nCol==BMS-1 ); 850 testcase( pTable->nCol==BMS-2 ); 851 for(i=0; i<mxBitCol; i++){ 852 if( extraCols & MASKBIT(i) ) nKeyCol++; 853 } 854 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 855 nKeyCol += pTable->nCol - BMS + 1; 856 } 857 858 /* Construct the Index object to describe this index */ 859 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); 860 if( pIdx==0 ) goto end_auto_index_create; 861 pLoop->u.btree.pIndex = pIdx; 862 pIdx->zName = "auto-index"; 863 pIdx->pTable = pTable; 864 n = 0; 865 idxCols = 0; 866 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 867 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 868 int iCol; 869 Bitmask cMask; 870 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 871 iCol = pTerm->u.x.leftColumn; 872 cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 873 testcase( iCol==BMS-1 ); 874 testcase( iCol==BMS ); 875 if( (idxCols & cMask)==0 ){ 876 Expr *pX = pTerm->pExpr; 877 idxCols |= cMask; 878 pIdx->aiColumn[n] = pTerm->u.x.leftColumn; 879 pColl = sqlite3ExprCompareCollSeq(pParse, pX); 880 assert( pColl!=0 || pParse->nErr>0 ); /* TH3 collate01.800 */ 881 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY; 882 n++; 883 } 884 } 885 } 886 assert( (u32)n==pLoop->u.btree.nEq ); 887 888 /* Add additional columns needed to make the automatic index into 889 ** a covering index */ 890 for(i=0; i<mxBitCol; i++){ 891 if( extraCols & MASKBIT(i) ){ 892 pIdx->aiColumn[n] = i; 893 pIdx->azColl[n] = sqlite3StrBINARY; 894 n++; 895 } 896 } 897 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 898 for(i=BMS-1; i<pTable->nCol; i++){ 899 pIdx->aiColumn[n] = i; 900 pIdx->azColl[n] = sqlite3StrBINARY; 901 n++; 902 } 903 } 904 assert( n==nKeyCol ); 905 pIdx->aiColumn[n] = XN_ROWID; 906 pIdx->azColl[n] = sqlite3StrBINARY; 907 908 /* Create the automatic index */ 909 assert( pLevel->iIdxCur>=0 ); 910 pLevel->iIdxCur = pParse->nTab++; 911 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); 912 sqlite3VdbeSetP4KeyInfo(pParse, pIdx); 913 VdbeComment((v, "for %s", pTable->zName)); 914 if( OptimizationEnabled(pParse->db, SQLITE_BloomFilter) ){ 915 pLevel->regFilter = ++pParse->nMem; 916 sqlite3VdbeAddOp2(v, OP_Blob, 10000, pLevel->regFilter); 917 } 918 919 /* Fill the automatic index with content */ 920 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom]; 921 if( pTabItem->fg.viaCoroutine ){ 922 int regYield = pTabItem->regReturn; 923 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0); 924 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); 925 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 926 VdbeCoverage(v); 927 VdbeComment((v, "next row of %s", pTabItem->pTab->zName)); 928 }else{ 929 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); 930 } 931 if( pPartial ){ 932 iContinue = sqlite3VdbeMakeLabel(pParse); 933 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); 934 pLoop->wsFlags |= WHERE_PARTIALIDX; 935 } 936 regRecord = sqlite3GetTempReg(pParse); 937 regBase = sqlite3GenerateIndexKey( 938 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0 939 ); 940 if( pLevel->regFilter ){ 941 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0, 942 regBase, pLoop->u.btree.nEq); 943 } 944 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 945 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 946 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); 947 if( pTabItem->fg.viaCoroutine ){ 948 sqlite3VdbeChangeP2(v, addrCounter, regBase+n); 949 testcase( pParse->db->mallocFailed ); 950 assert( pLevel->iIdxCur>0 ); 951 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur, 952 pTabItem->regResult, pLevel->iIdxCur); 953 sqlite3VdbeGoto(v, addrTop); 954 pTabItem->fg.viaCoroutine = 0; 955 }else{ 956 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); 957 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 958 } 959 sqlite3VdbeJumpHere(v, addrTop); 960 sqlite3ReleaseTempReg(pParse, regRecord); 961 962 /* Jump here when skipping the initialization */ 963 sqlite3VdbeJumpHere(v, addrInit); 964 965 end_auto_index_create: 966 sqlite3ExprDelete(pParse->db, pPartial); 967 } 968 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 969 970 /* 971 ** Generate bytecode that will initialize a Bloom filter that is appropriate 972 ** for pLevel. 973 ** 974 ** If there are inner loops within pLevel that have the WHERE_BLOOMFILTER 975 ** flag set, initialize a Bloomfilter for them as well. Except don't do 976 ** this recursive initialization if the SQLITE_BloomPulldown optimization has 977 ** been turned off. 978 ** 979 ** When the Bloom filter is initialized, the WHERE_BLOOMFILTER flag is cleared 980 ** from the loop, but the regFilter value is set to a register that implements 981 ** the Bloom filter. When regFilter is positive, the 982 ** sqlite3WhereCodeOneLoopStart() will generate code to test the Bloom filter 983 ** and skip the subsequence B-Tree seek if the Bloom filter indicates that 984 ** no matching rows exist. 985 ** 986 ** This routine may only be called if it has previously been determined that 987 ** the loop would benefit from a Bloom filter, and the WHERE_BLOOMFILTER bit 988 ** is set. 989 */ 990 static SQLITE_NOINLINE void sqlite3ConstructBloomFilter( 991 WhereInfo *pWInfo, /* The WHERE clause */ 992 int iLevel, /* Index in pWInfo->a[] that is pLevel */ 993 WhereLevel *pLevel, /* Make a Bloom filter for this FROM term */ 994 Bitmask notReady /* Loops that are not ready */ 995 ){ 996 int addrOnce; /* Address of opening OP_Once */ 997 int addrTop; /* Address of OP_Rewind */ 998 int addrCont; /* Jump here to skip a row */ 999 const WhereTerm *pTerm; /* For looping over WHERE clause terms */ 1000 const WhereTerm *pWCEnd; /* Last WHERE clause term */ 1001 Parse *pParse = pWInfo->pParse; /* Parsing context */ 1002 Vdbe *v = pParse->pVdbe; /* VDBE under construction */ 1003 WhereLoop *pLoop = pLevel->pWLoop; /* The loop being coded */ 1004 int iCur; /* Cursor for table getting the filter */ 1005 1006 assert( pLoop!=0 ); 1007 assert( v!=0 ); 1008 assert( pLoop->wsFlags & WHERE_BLOOMFILTER ); 1009 1010 addrOnce = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 1011 do{ 1012 const SrcItem *pItem; 1013 const Table *pTab; 1014 u64 sz; 1015 sqlite3WhereExplainBloomFilter(pParse, pWInfo, pLevel); 1016 addrCont = sqlite3VdbeMakeLabel(pParse); 1017 iCur = pLevel->iTabCur; 1018 pLevel->regFilter = ++pParse->nMem; 1019 1020 /* The Bloom filter is a Blob held in a register. Initialize it 1021 ** to zero-filled blob of at least 80K bits, but maybe more if the 1022 ** estimated size of the table is larger. We could actually 1023 ** measure the size of the table at run-time using OP_Count with 1024 ** P3==1 and use that value to initialize the blob. But that makes 1025 ** testing complicated. By basing the blob size on the value in the 1026 ** sqlite_stat1 table, testing is much easier. 1027 */ 1028 pItem = &pWInfo->pTabList->a[pLevel->iFrom]; 1029 assert( pItem!=0 ); 1030 pTab = pItem->pTab; 1031 assert( pTab!=0 ); 1032 sz = sqlite3LogEstToInt(pTab->nRowLogEst); 1033 if( sz<10000 ){ 1034 sz = 10000; 1035 }else if( sz>10000000 ){ 1036 sz = 10000000; 1037 } 1038 sqlite3VdbeAddOp2(v, OP_Blob, (int)sz, pLevel->regFilter); 1039 1040 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, iCur); VdbeCoverage(v); 1041 pWCEnd = &pWInfo->sWC.a[pWInfo->sWC.nTerm]; 1042 for(pTerm=pWInfo->sWC.a; pTerm<pWCEnd; pTerm++){ 1043 Expr *pExpr = pTerm->pExpr; 1044 if( (pTerm->wtFlags & TERM_VIRTUAL)==0 1045 && sqlite3ExprIsTableConstant(pExpr, iCur) 1046 ){ 1047 sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL); 1048 } 1049 } 1050 if( pLoop->wsFlags & WHERE_IPK ){ 1051 int r1 = sqlite3GetTempReg(pParse); 1052 sqlite3VdbeAddOp2(v, OP_Rowid, iCur, r1); 1053 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0, r1, 1); 1054 sqlite3ReleaseTempReg(pParse, r1); 1055 }else{ 1056 Index *pIdx = pLoop->u.btree.pIndex; 1057 int n = pLoop->u.btree.nEq; 1058 int r1 = sqlite3GetTempRange(pParse, n); 1059 int jj; 1060 for(jj=0; jj<n; jj++){ 1061 int iCol = pIdx->aiColumn[jj]; 1062 assert( pIdx->pTable==pItem->pTab ); 1063 sqlite3ExprCodeGetColumnOfTable(v, pIdx->pTable, iCur, iCol,r1+jj); 1064 } 1065 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0, r1, n); 1066 sqlite3ReleaseTempRange(pParse, r1, n); 1067 } 1068 sqlite3VdbeResolveLabel(v, addrCont); 1069 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); 1070 VdbeCoverage(v); 1071 sqlite3VdbeJumpHere(v, addrTop); 1072 pLoop->wsFlags &= ~WHERE_BLOOMFILTER; 1073 if( OptimizationDisabled(pParse->db, SQLITE_BloomPulldown) ) break; 1074 while( ++iLevel < pWInfo->nLevel ){ 1075 pLevel = &pWInfo->a[iLevel]; 1076 pLoop = pLevel->pWLoop; 1077 if( NEVER(pLoop==0) ) continue; 1078 if( pLoop->prereq & notReady ) continue; 1079 if( (pLoop->wsFlags & (WHERE_BLOOMFILTER|WHERE_COLUMN_IN)) 1080 ==WHERE_BLOOMFILTER 1081 ){ 1082 /* This is a candidate for bloom-filter pull-down (early evaluation). 1083 ** The test that WHERE_COLUMN_IN is omitted is important, as we are 1084 ** not able to do early evaluation of bloom filters that make use of 1085 ** the IN operator */ 1086 break; 1087 } 1088 } 1089 }while( iLevel < pWInfo->nLevel ); 1090 sqlite3VdbeJumpHere(v, addrOnce); 1091 } 1092 1093 1094 #ifndef SQLITE_OMIT_VIRTUALTABLE 1095 /* 1096 ** Allocate and populate an sqlite3_index_info structure. It is the 1097 ** responsibility of the caller to eventually release the structure 1098 ** by passing the pointer returned by this function to sqlite3_free(). 1099 */ 1100 static sqlite3_index_info *allocateIndexInfo( 1101 Parse *pParse, /* The parsing context */ 1102 WhereClause *pWC, /* The WHERE clause being analyzed */ 1103 Bitmask mUnusable, /* Ignore terms with these prereqs */ 1104 SrcItem *pSrc, /* The FROM clause term that is the vtab */ 1105 ExprList *pOrderBy, /* The ORDER BY clause */ 1106 u16 *pmNoOmit /* Mask of terms not to omit */ 1107 ){ 1108 int i, j; 1109 int nTerm; 1110 struct sqlite3_index_constraint *pIdxCons; 1111 struct sqlite3_index_orderby *pIdxOrderBy; 1112 struct sqlite3_index_constraint_usage *pUsage; 1113 struct HiddenIndexInfo *pHidden; 1114 WhereTerm *pTerm; 1115 int nOrderBy; 1116 sqlite3_index_info *pIdxInfo; 1117 u16 mNoOmit = 0; 1118 const Table *pTab; 1119 1120 assert( pSrc!=0 ); 1121 pTab = pSrc->pTab; 1122 assert( pTab!=0 ); 1123 assert( IsVirtual(pTab) ); 1124 1125 /* Find all WHERE clause constraints referring to this virtual table. 1126 ** Mark each term with the TERM_OK flag. Set nTerm to the number of 1127 ** terms found. 1128 */ 1129 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 1130 pTerm->wtFlags &= ~TERM_OK; 1131 if( pTerm->leftCursor != pSrc->iCursor ) continue; 1132 if( pTerm->prereqRight & mUnusable ) continue; 1133 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 1134 testcase( pTerm->eOperator & WO_IN ); 1135 testcase( pTerm->eOperator & WO_ISNULL ); 1136 testcase( pTerm->eOperator & WO_IS ); 1137 testcase( pTerm->eOperator & WO_ALL ); 1138 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 1139 if( pTerm->wtFlags & TERM_VNULL ) continue; 1140 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 1141 assert( pTerm->u.x.leftColumn>=XN_ROWID ); 1142 assert( pTerm->u.x.leftColumn<pTab->nCol ); 1143 1144 /* tag-20191211-002: WHERE-clause constraints are not useful to the 1145 ** right-hand table of a LEFT JOIN. See tag-20191211-001 for the 1146 ** equivalent restriction for ordinary tables. */ 1147 if( (pSrc->fg.jointype & JT_LEFT)!=0 1148 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 1149 ){ 1150 continue; 1151 } 1152 nTerm++; 1153 pTerm->wtFlags |= TERM_OK; 1154 } 1155 1156 /* If the ORDER BY clause contains only columns in the current 1157 ** virtual table then allocate space for the aOrderBy part of 1158 ** the sqlite3_index_info structure. 1159 */ 1160 nOrderBy = 0; 1161 if( pOrderBy ){ 1162 int n = pOrderBy->nExpr; 1163 for(i=0; i<n; i++){ 1164 Expr *pExpr = pOrderBy->a[i].pExpr; 1165 Expr *pE2; 1166 1167 /* Skip over constant terms in the ORDER BY clause */ 1168 if( sqlite3ExprIsConstant(pExpr) ){ 1169 continue; 1170 } 1171 1172 /* Virtual tables are unable to deal with NULLS FIRST */ 1173 if( pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL ) break; 1174 1175 /* First case - a direct column references without a COLLATE operator */ 1176 if( pExpr->op==TK_COLUMN && pExpr->iTable==pSrc->iCursor ){ 1177 assert( pExpr->iColumn>=XN_ROWID && pExpr->iColumn<pTab->nCol ); 1178 continue; 1179 } 1180 1181 /* 2nd case - a column reference with a COLLATE operator. Only match 1182 ** of the COLLATE operator matches the collation of the column. */ 1183 if( pExpr->op==TK_COLLATE 1184 && (pE2 = pExpr->pLeft)->op==TK_COLUMN 1185 && pE2->iTable==pSrc->iCursor 1186 ){ 1187 const char *zColl; /* The collating sequence name */ 1188 assert( !ExprHasProperty(pExpr, EP_IntValue) ); 1189 assert( pExpr->u.zToken!=0 ); 1190 assert( pE2->iColumn>=XN_ROWID && pE2->iColumn<pTab->nCol ); 1191 pExpr->iColumn = pE2->iColumn; 1192 if( pE2->iColumn<0 ) continue; /* Collseq does not matter for rowid */ 1193 zColl = sqlite3ColumnColl(&pTab->aCol[pE2->iColumn]); 1194 if( zColl==0 ) zColl = sqlite3StrBINARY; 1195 if( sqlite3_stricmp(pExpr->u.zToken, zColl)==0 ) continue; 1196 } 1197 1198 /* No matches cause a break out of the loop */ 1199 break; 1200 } 1201 if( i==n){ 1202 nOrderBy = n; 1203 } 1204 } 1205 1206 /* Allocate the sqlite3_index_info structure 1207 */ 1208 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 1209 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 1210 + sizeof(*pIdxOrderBy)*nOrderBy + sizeof(*pHidden) ); 1211 if( pIdxInfo==0 ){ 1212 sqlite3ErrorMsg(pParse, "out of memory"); 1213 return 0; 1214 } 1215 pHidden = (struct HiddenIndexInfo*)&pIdxInfo[1]; 1216 pIdxCons = (struct sqlite3_index_constraint*)&pHidden[1]; 1217 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 1218 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 1219 pIdxInfo->aConstraint = pIdxCons; 1220 pIdxInfo->aOrderBy = pIdxOrderBy; 1221 pIdxInfo->aConstraintUsage = pUsage; 1222 pHidden->pWC = pWC; 1223 pHidden->pParse = pParse; 1224 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 1225 u16 op; 1226 if( (pTerm->wtFlags & TERM_OK)==0 ) continue; 1227 pIdxCons[j].iColumn = pTerm->u.x.leftColumn; 1228 pIdxCons[j].iTermOffset = i; 1229 op = pTerm->eOperator & WO_ALL; 1230 if( op==WO_IN ) op = WO_EQ; 1231 if( op==WO_AUX ){ 1232 pIdxCons[j].op = pTerm->eMatchOp; 1233 }else if( op & (WO_ISNULL|WO_IS) ){ 1234 if( op==WO_ISNULL ){ 1235 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_ISNULL; 1236 }else{ 1237 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_IS; 1238 } 1239 }else{ 1240 pIdxCons[j].op = (u8)op; 1241 /* The direct assignment in the previous line is possible only because 1242 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 1243 ** following asserts verify this fact. */ 1244 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 1245 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 1246 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 1247 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 1248 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 1249 assert( pTerm->eOperator&(WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_AUX) ); 1250 1251 if( op & (WO_LT|WO_LE|WO_GT|WO_GE) 1252 && sqlite3ExprIsVector(pTerm->pExpr->pRight) 1253 ){ 1254 testcase( j!=i ); 1255 if( j<16 ) mNoOmit |= (1 << j); 1256 if( op==WO_LT ) pIdxCons[j].op = WO_LE; 1257 if( op==WO_GT ) pIdxCons[j].op = WO_GE; 1258 } 1259 } 1260 1261 j++; 1262 } 1263 assert( j==nTerm ); 1264 pIdxInfo->nConstraint = j; 1265 for(i=j=0; i<nOrderBy; i++){ 1266 Expr *pExpr = pOrderBy->a[i].pExpr; 1267 if( sqlite3ExprIsConstant(pExpr) ) continue; 1268 assert( pExpr->op==TK_COLUMN 1269 || (pExpr->op==TK_COLLATE && pExpr->pLeft->op==TK_COLUMN 1270 && pExpr->iColumn==pExpr->pLeft->iColumn) ); 1271 pIdxOrderBy[j].iColumn = pExpr->iColumn; 1272 pIdxOrderBy[j].desc = pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC; 1273 j++; 1274 } 1275 pIdxInfo->nOrderBy = j; 1276 1277 *pmNoOmit = mNoOmit; 1278 return pIdxInfo; 1279 } 1280 1281 /* 1282 ** The table object reference passed as the second argument to this function 1283 ** must represent a virtual table. This function invokes the xBestIndex() 1284 ** method of the virtual table with the sqlite3_index_info object that 1285 ** comes in as the 3rd argument to this function. 1286 ** 1287 ** If an error occurs, pParse is populated with an error message and an 1288 ** appropriate error code is returned. A return of SQLITE_CONSTRAINT from 1289 ** xBestIndex is not considered an error. SQLITE_CONSTRAINT indicates that 1290 ** the current configuration of "unusable" flags in sqlite3_index_info can 1291 ** not result in a valid plan. 1292 ** 1293 ** Whether or not an error is returned, it is the responsibility of the 1294 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 1295 ** that this is required. 1296 */ 1297 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 1298 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 1299 int rc; 1300 1301 whereTraceIndexInfoInputs(p); 1302 rc = pVtab->pModule->xBestIndex(pVtab, p); 1303 whereTraceIndexInfoOutputs(p); 1304 1305 if( rc!=SQLITE_OK && rc!=SQLITE_CONSTRAINT ){ 1306 if( rc==SQLITE_NOMEM ){ 1307 sqlite3OomFault(pParse->db); 1308 }else if( !pVtab->zErrMsg ){ 1309 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 1310 }else{ 1311 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 1312 } 1313 } 1314 sqlite3_free(pVtab->zErrMsg); 1315 pVtab->zErrMsg = 0; 1316 return rc; 1317 } 1318 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 1319 1320 #ifdef SQLITE_ENABLE_STAT4 1321 /* 1322 ** Estimate the location of a particular key among all keys in an 1323 ** index. Store the results in aStat as follows: 1324 ** 1325 ** aStat[0] Est. number of rows less than pRec 1326 ** aStat[1] Est. number of rows equal to pRec 1327 ** 1328 ** Return the index of the sample that is the smallest sample that 1329 ** is greater than or equal to pRec. Note that this index is not an index 1330 ** into the aSample[] array - it is an index into a virtual set of samples 1331 ** based on the contents of aSample[] and the number of fields in record 1332 ** pRec. 1333 */ 1334 static int whereKeyStats( 1335 Parse *pParse, /* Database connection */ 1336 Index *pIdx, /* Index to consider domain of */ 1337 UnpackedRecord *pRec, /* Vector of values to consider */ 1338 int roundUp, /* Round up if true. Round down if false */ 1339 tRowcnt *aStat /* OUT: stats written here */ 1340 ){ 1341 IndexSample *aSample = pIdx->aSample; 1342 int iCol; /* Index of required stats in anEq[] etc. */ 1343 int i; /* Index of first sample >= pRec */ 1344 int iSample; /* Smallest sample larger than or equal to pRec */ 1345 int iMin = 0; /* Smallest sample not yet tested */ 1346 int iTest; /* Next sample to test */ 1347 int res; /* Result of comparison operation */ 1348 int nField; /* Number of fields in pRec */ 1349 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 1350 1351 #ifndef SQLITE_DEBUG 1352 UNUSED_PARAMETER( pParse ); 1353 #endif 1354 assert( pRec!=0 ); 1355 assert( pIdx->nSample>0 ); 1356 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1357 1358 /* Do a binary search to find the first sample greater than or equal 1359 ** to pRec. If pRec contains a single field, the set of samples to search 1360 ** is simply the aSample[] array. If the samples in aSample[] contain more 1361 ** than one fields, all fields following the first are ignored. 1362 ** 1363 ** If pRec contains N fields, where N is more than one, then as well as the 1364 ** samples in aSample[] (truncated to N fields), the search also has to 1365 ** consider prefixes of those samples. For example, if the set of samples 1366 ** in aSample is: 1367 ** 1368 ** aSample[0] = (a, 5) 1369 ** aSample[1] = (a, 10) 1370 ** aSample[2] = (b, 5) 1371 ** aSample[3] = (c, 100) 1372 ** aSample[4] = (c, 105) 1373 ** 1374 ** Then the search space should ideally be the samples above and the 1375 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1376 ** the code actually searches this set: 1377 ** 1378 ** 0: (a) 1379 ** 1: (a, 5) 1380 ** 2: (a, 10) 1381 ** 3: (a, 10) 1382 ** 4: (b) 1383 ** 5: (b, 5) 1384 ** 6: (c) 1385 ** 7: (c, 100) 1386 ** 8: (c, 105) 1387 ** 9: (c, 105) 1388 ** 1389 ** For each sample in the aSample[] array, N samples are present in the 1390 ** effective sample array. In the above, samples 0 and 1 are based on 1391 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1392 ** 1393 ** Often, sample i of each block of N effective samples has (i+1) fields. 1394 ** Except, each sample may be extended to ensure that it is greater than or 1395 ** equal to the previous sample in the array. For example, in the above, 1396 ** sample 2 is the first sample of a block of N samples, so at first it 1397 ** appears that it should be 1 field in size. However, that would make it 1398 ** smaller than sample 1, so the binary search would not work. As a result, 1399 ** it is extended to two fields. The duplicates that this creates do not 1400 ** cause any problems. 1401 */ 1402 nField = pRec->nField; 1403 iCol = 0; 1404 iSample = pIdx->nSample * nField; 1405 do{ 1406 int iSamp; /* Index in aSample[] of test sample */ 1407 int n; /* Number of fields in test sample */ 1408 1409 iTest = (iMin+iSample)/2; 1410 iSamp = iTest / nField; 1411 if( iSamp>0 ){ 1412 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1413 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1414 ** fields that is greater than the previous effective sample. */ 1415 for(n=(iTest % nField) + 1; n<nField; n++){ 1416 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1417 } 1418 }else{ 1419 n = iTest + 1; 1420 } 1421 1422 pRec->nField = n; 1423 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1424 if( res<0 ){ 1425 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1426 iMin = iTest+1; 1427 }else if( res==0 && n<nField ){ 1428 iLower = aSample[iSamp].anLt[n-1]; 1429 iMin = iTest+1; 1430 res = -1; 1431 }else{ 1432 iSample = iTest; 1433 iCol = n-1; 1434 } 1435 }while( res && iMin<iSample ); 1436 i = iSample / nField; 1437 1438 #ifdef SQLITE_DEBUG 1439 /* The following assert statements check that the binary search code 1440 ** above found the right answer. This block serves no purpose other 1441 ** than to invoke the asserts. */ 1442 if( pParse->db->mallocFailed==0 ){ 1443 if( res==0 ){ 1444 /* If (res==0) is true, then pRec must be equal to sample i. */ 1445 assert( i<pIdx->nSample ); 1446 assert( iCol==nField-1 ); 1447 pRec->nField = nField; 1448 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1449 || pParse->db->mallocFailed 1450 ); 1451 }else{ 1452 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1453 ** all samples in the aSample[] array, pRec must be smaller than the 1454 ** (iCol+1) field prefix of sample i. */ 1455 assert( i<=pIdx->nSample && i>=0 ); 1456 pRec->nField = iCol+1; 1457 assert( i==pIdx->nSample 1458 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1459 || pParse->db->mallocFailed ); 1460 1461 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1462 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1463 ** be greater than or equal to the (iCol) field prefix of sample i. 1464 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1465 if( iCol>0 ){ 1466 pRec->nField = iCol; 1467 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1468 || pParse->db->mallocFailed ); 1469 } 1470 if( i>0 ){ 1471 pRec->nField = nField; 1472 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1473 || pParse->db->mallocFailed ); 1474 } 1475 } 1476 } 1477 #endif /* ifdef SQLITE_DEBUG */ 1478 1479 if( res==0 ){ 1480 /* Record pRec is equal to sample i */ 1481 assert( iCol==nField-1 ); 1482 aStat[0] = aSample[i].anLt[iCol]; 1483 aStat[1] = aSample[i].anEq[iCol]; 1484 }else{ 1485 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1486 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1487 ** is larger than all samples in the array. */ 1488 tRowcnt iUpper, iGap; 1489 if( i>=pIdx->nSample ){ 1490 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1491 }else{ 1492 iUpper = aSample[i].anLt[iCol]; 1493 } 1494 1495 if( iLower>=iUpper ){ 1496 iGap = 0; 1497 }else{ 1498 iGap = iUpper - iLower; 1499 } 1500 if( roundUp ){ 1501 iGap = (iGap*2)/3; 1502 }else{ 1503 iGap = iGap/3; 1504 } 1505 aStat[0] = iLower + iGap; 1506 aStat[1] = pIdx->aAvgEq[nField-1]; 1507 } 1508 1509 /* Restore the pRec->nField value before returning. */ 1510 pRec->nField = nField; 1511 return i; 1512 } 1513 #endif /* SQLITE_ENABLE_STAT4 */ 1514 1515 /* 1516 ** If it is not NULL, pTerm is a term that provides an upper or lower 1517 ** bound on a range scan. Without considering pTerm, it is estimated 1518 ** that the scan will visit nNew rows. This function returns the number 1519 ** estimated to be visited after taking pTerm into account. 1520 ** 1521 ** If the user explicitly specified a likelihood() value for this term, 1522 ** then the return value is the likelihood multiplied by the number of 1523 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1524 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1525 */ 1526 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1527 LogEst nRet = nNew; 1528 if( pTerm ){ 1529 if( pTerm->truthProb<=0 ){ 1530 nRet += pTerm->truthProb; 1531 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1532 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1533 } 1534 } 1535 return nRet; 1536 } 1537 1538 1539 #ifdef SQLITE_ENABLE_STAT4 1540 /* 1541 ** Return the affinity for a single column of an index. 1542 */ 1543 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1544 assert( iCol>=0 && iCol<pIdx->nColumn ); 1545 if( !pIdx->zColAff ){ 1546 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1547 } 1548 assert( pIdx->zColAff[iCol]!=0 ); 1549 return pIdx->zColAff[iCol]; 1550 } 1551 #endif 1552 1553 1554 #ifdef SQLITE_ENABLE_STAT4 1555 /* 1556 ** This function is called to estimate the number of rows visited by a 1557 ** range-scan on a skip-scan index. For example: 1558 ** 1559 ** CREATE INDEX i1 ON t1(a, b, c); 1560 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1561 ** 1562 ** Value pLoop->nOut is currently set to the estimated number of rows 1563 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1564 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1565 ** on the stat4 data for the index. this scan will be peformed multiple 1566 ** times (once for each (a,b) combination that matches a=?) is dealt with 1567 ** by the caller. 1568 ** 1569 ** It does this by scanning through all stat4 samples, comparing values 1570 ** extracted from pLower and pUpper with the corresponding column in each 1571 ** sample. If L and U are the number of samples found to be less than or 1572 ** equal to the values extracted from pLower and pUpper respectively, and 1573 ** N is the total number of samples, the pLoop->nOut value is adjusted 1574 ** as follows: 1575 ** 1576 ** nOut = nOut * ( min(U - L, 1) / N ) 1577 ** 1578 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1579 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1580 ** U is set to N. 1581 ** 1582 ** Normally, this function sets *pbDone to 1 before returning. However, 1583 ** if no value can be extracted from either pLower or pUpper (and so the 1584 ** estimate of the number of rows delivered remains unchanged), *pbDone 1585 ** is left as is. 1586 ** 1587 ** If an error occurs, an SQLite error code is returned. Otherwise, 1588 ** SQLITE_OK. 1589 */ 1590 static int whereRangeSkipScanEst( 1591 Parse *pParse, /* Parsing & code generating context */ 1592 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1593 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1594 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1595 int *pbDone /* Set to true if at least one expr. value extracted */ 1596 ){ 1597 Index *p = pLoop->u.btree.pIndex; 1598 int nEq = pLoop->u.btree.nEq; 1599 sqlite3 *db = pParse->db; 1600 int nLower = -1; 1601 int nUpper = p->nSample+1; 1602 int rc = SQLITE_OK; 1603 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1604 CollSeq *pColl; 1605 1606 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1607 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1608 sqlite3_value *pVal = 0; /* Value extracted from record */ 1609 1610 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1611 if( pLower ){ 1612 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1613 nLower = 0; 1614 } 1615 if( pUpper && rc==SQLITE_OK ){ 1616 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1617 nUpper = p2 ? 0 : p->nSample; 1618 } 1619 1620 if( p1 || p2 ){ 1621 int i; 1622 int nDiff; 1623 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1624 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1625 if( rc==SQLITE_OK && p1 ){ 1626 int res = sqlite3MemCompare(p1, pVal, pColl); 1627 if( res>=0 ) nLower++; 1628 } 1629 if( rc==SQLITE_OK && p2 ){ 1630 int res = sqlite3MemCompare(p2, pVal, pColl); 1631 if( res>=0 ) nUpper++; 1632 } 1633 } 1634 nDiff = (nUpper - nLower); 1635 if( nDiff<=0 ) nDiff = 1; 1636 1637 /* If there is both an upper and lower bound specified, and the 1638 ** comparisons indicate that they are close together, use the fallback 1639 ** method (assume that the scan visits 1/64 of the rows) for estimating 1640 ** the number of rows visited. Otherwise, estimate the number of rows 1641 ** using the method described in the header comment for this function. */ 1642 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1643 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1644 pLoop->nOut -= nAdjust; 1645 *pbDone = 1; 1646 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1647 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1648 } 1649 1650 }else{ 1651 assert( *pbDone==0 ); 1652 } 1653 1654 sqlite3ValueFree(p1); 1655 sqlite3ValueFree(p2); 1656 sqlite3ValueFree(pVal); 1657 1658 return rc; 1659 } 1660 #endif /* SQLITE_ENABLE_STAT4 */ 1661 1662 /* 1663 ** This function is used to estimate the number of rows that will be visited 1664 ** by scanning an index for a range of values. The range may have an upper 1665 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1666 ** and lower bounds are represented by pLower and pUpper respectively. For 1667 ** example, assuming that index p is on t1(a): 1668 ** 1669 ** ... FROM t1 WHERE a > ? AND a < ? ... 1670 ** |_____| |_____| 1671 ** | | 1672 ** pLower pUpper 1673 ** 1674 ** If either of the upper or lower bound is not present, then NULL is passed in 1675 ** place of the corresponding WhereTerm. 1676 ** 1677 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1678 ** column subject to the range constraint. Or, equivalently, the number of 1679 ** equality constraints optimized by the proposed index scan. For example, 1680 ** assuming index p is on t1(a, b), and the SQL query is: 1681 ** 1682 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1683 ** 1684 ** then nEq is set to 1 (as the range restricted column, b, is the second 1685 ** left-most column of the index). Or, if the query is: 1686 ** 1687 ** ... FROM t1 WHERE a > ? AND a < ? ... 1688 ** 1689 ** then nEq is set to 0. 1690 ** 1691 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1692 ** number of rows that the index scan is expected to visit without 1693 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1694 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1695 ** to account for the range constraints pLower and pUpper. 1696 ** 1697 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1698 ** used, a single range inequality reduces the search space by a factor of 4. 1699 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1700 ** rows visited by a factor of 64. 1701 */ 1702 static int whereRangeScanEst( 1703 Parse *pParse, /* Parsing & code generating context */ 1704 WhereLoopBuilder *pBuilder, 1705 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1706 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1707 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1708 ){ 1709 int rc = SQLITE_OK; 1710 int nOut = pLoop->nOut; 1711 LogEst nNew; 1712 1713 #ifdef SQLITE_ENABLE_STAT4 1714 Index *p = pLoop->u.btree.pIndex; 1715 int nEq = pLoop->u.btree.nEq; 1716 1717 if( p->nSample>0 && ALWAYS(nEq<p->nSampleCol) 1718 && OptimizationEnabled(pParse->db, SQLITE_Stat4) 1719 ){ 1720 if( nEq==pBuilder->nRecValid ){ 1721 UnpackedRecord *pRec = pBuilder->pRec; 1722 tRowcnt a[2]; 1723 int nBtm = pLoop->u.btree.nBtm; 1724 int nTop = pLoop->u.btree.nTop; 1725 1726 /* Variable iLower will be set to the estimate of the number of rows in 1727 ** the index that are less than the lower bound of the range query. The 1728 ** lower bound being the concatenation of $P and $L, where $P is the 1729 ** key-prefix formed by the nEq values matched against the nEq left-most 1730 ** columns of the index, and $L is the value in pLower. 1731 ** 1732 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1733 ** is not a simple variable or literal value), the lower bound of the 1734 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1735 ** if $L is available, whereKeyStats() is called for both ($P) and 1736 ** ($P:$L) and the larger of the two returned values is used. 1737 ** 1738 ** Similarly, iUpper is to be set to the estimate of the number of rows 1739 ** less than the upper bound of the range query. Where the upper bound 1740 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1741 ** of iUpper are requested of whereKeyStats() and the smaller used. 1742 ** 1743 ** The number of rows between the two bounds is then just iUpper-iLower. 1744 */ 1745 tRowcnt iLower; /* Rows less than the lower bound */ 1746 tRowcnt iUpper; /* Rows less than the upper bound */ 1747 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1748 int iUprIdx = -1; /* aSample[] for the upper bound */ 1749 1750 if( pRec ){ 1751 testcase( pRec->nField!=pBuilder->nRecValid ); 1752 pRec->nField = pBuilder->nRecValid; 1753 } 1754 /* Determine iLower and iUpper using ($P) only. */ 1755 if( nEq==0 ){ 1756 iLower = 0; 1757 iUpper = p->nRowEst0; 1758 }else{ 1759 /* Note: this call could be optimized away - since the same values must 1760 ** have been requested when testing key $P in whereEqualScanEst(). */ 1761 whereKeyStats(pParse, p, pRec, 0, a); 1762 iLower = a[0]; 1763 iUpper = a[0] + a[1]; 1764 } 1765 1766 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1767 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1768 assert( p->aSortOrder!=0 ); 1769 if( p->aSortOrder[nEq] ){ 1770 /* The roles of pLower and pUpper are swapped for a DESC index */ 1771 SWAP(WhereTerm*, pLower, pUpper); 1772 SWAP(int, nBtm, nTop); 1773 } 1774 1775 /* If possible, improve on the iLower estimate using ($P:$L). */ 1776 if( pLower ){ 1777 int n; /* Values extracted from pExpr */ 1778 Expr *pExpr = pLower->pExpr->pRight; 1779 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1780 if( rc==SQLITE_OK && n ){ 1781 tRowcnt iNew; 1782 u16 mask = WO_GT|WO_LE; 1783 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1784 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1785 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1786 if( iNew>iLower ) iLower = iNew; 1787 nOut--; 1788 pLower = 0; 1789 } 1790 } 1791 1792 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1793 if( pUpper ){ 1794 int n; /* Values extracted from pExpr */ 1795 Expr *pExpr = pUpper->pExpr->pRight; 1796 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1797 if( rc==SQLITE_OK && n ){ 1798 tRowcnt iNew; 1799 u16 mask = WO_GT|WO_LE; 1800 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1801 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1802 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1803 if( iNew<iUpper ) iUpper = iNew; 1804 nOut--; 1805 pUpper = 0; 1806 } 1807 } 1808 1809 pBuilder->pRec = pRec; 1810 if( rc==SQLITE_OK ){ 1811 if( iUpper>iLower ){ 1812 nNew = sqlite3LogEst(iUpper - iLower); 1813 /* TUNING: If both iUpper and iLower are derived from the same 1814 ** sample, then assume they are 4x more selective. This brings 1815 ** the estimated selectivity more in line with what it would be 1816 ** if estimated without the use of STAT4 tables. */ 1817 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1818 }else{ 1819 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1820 } 1821 if( nNew<nOut ){ 1822 nOut = nNew; 1823 } 1824 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1825 (u32)iLower, (u32)iUpper, nOut)); 1826 } 1827 }else{ 1828 int bDone = 0; 1829 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1830 if( bDone ) return rc; 1831 } 1832 } 1833 #else 1834 UNUSED_PARAMETER(pParse); 1835 UNUSED_PARAMETER(pBuilder); 1836 assert( pLower || pUpper ); 1837 #endif 1838 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1839 nNew = whereRangeAdjust(pLower, nOut); 1840 nNew = whereRangeAdjust(pUpper, nNew); 1841 1842 /* TUNING: If there is both an upper and lower limit and neither limit 1843 ** has an application-defined likelihood(), assume the range is 1844 ** reduced by an additional 75%. This means that, by default, an open-ended 1845 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1846 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1847 ** match 1/64 of the index. */ 1848 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1849 nNew -= 20; 1850 } 1851 1852 nOut -= (pLower!=0) + (pUpper!=0); 1853 if( nNew<10 ) nNew = 10; 1854 if( nNew<nOut ) nOut = nNew; 1855 #if defined(WHERETRACE_ENABLED) 1856 if( pLoop->nOut>nOut ){ 1857 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1858 pLoop->nOut, nOut)); 1859 } 1860 #endif 1861 pLoop->nOut = (LogEst)nOut; 1862 return rc; 1863 } 1864 1865 #ifdef SQLITE_ENABLE_STAT4 1866 /* 1867 ** Estimate the number of rows that will be returned based on 1868 ** an equality constraint x=VALUE and where that VALUE occurs in 1869 ** the histogram data. This only works when x is the left-most 1870 ** column of an index and sqlite_stat4 histogram data is available 1871 ** for that index. When pExpr==NULL that means the constraint is 1872 ** "x IS NULL" instead of "x=VALUE". 1873 ** 1874 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1875 ** If unable to make an estimate, leave *pnRow unchanged and return 1876 ** non-zero. 1877 ** 1878 ** This routine can fail if it is unable to load a collating sequence 1879 ** required for string comparison, or if unable to allocate memory 1880 ** for a UTF conversion required for comparison. The error is stored 1881 ** in the pParse structure. 1882 */ 1883 static int whereEqualScanEst( 1884 Parse *pParse, /* Parsing & code generating context */ 1885 WhereLoopBuilder *pBuilder, 1886 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1887 tRowcnt *pnRow /* Write the revised row estimate here */ 1888 ){ 1889 Index *p = pBuilder->pNew->u.btree.pIndex; 1890 int nEq = pBuilder->pNew->u.btree.nEq; 1891 UnpackedRecord *pRec = pBuilder->pRec; 1892 int rc; /* Subfunction return code */ 1893 tRowcnt a[2]; /* Statistics */ 1894 int bOk; 1895 1896 assert( nEq>=1 ); 1897 assert( nEq<=p->nColumn ); 1898 assert( p->aSample!=0 ); 1899 assert( p->nSample>0 ); 1900 assert( pBuilder->nRecValid<nEq ); 1901 1902 /* If values are not available for all fields of the index to the left 1903 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1904 if( pBuilder->nRecValid<(nEq-1) ){ 1905 return SQLITE_NOTFOUND; 1906 } 1907 1908 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1909 ** below would return the same value. */ 1910 if( nEq>=p->nColumn ){ 1911 *pnRow = 1; 1912 return SQLITE_OK; 1913 } 1914 1915 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1916 pBuilder->pRec = pRec; 1917 if( rc!=SQLITE_OK ) return rc; 1918 if( bOk==0 ) return SQLITE_NOTFOUND; 1919 pBuilder->nRecValid = nEq; 1920 1921 whereKeyStats(pParse, p, pRec, 0, a); 1922 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1923 p->zName, nEq-1, (int)a[1])); 1924 *pnRow = a[1]; 1925 1926 return rc; 1927 } 1928 #endif /* SQLITE_ENABLE_STAT4 */ 1929 1930 #ifdef SQLITE_ENABLE_STAT4 1931 /* 1932 ** Estimate the number of rows that will be returned based on 1933 ** an IN constraint where the right-hand side of the IN operator 1934 ** is a list of values. Example: 1935 ** 1936 ** WHERE x IN (1,2,3,4) 1937 ** 1938 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1939 ** If unable to make an estimate, leave *pnRow unchanged and return 1940 ** non-zero. 1941 ** 1942 ** This routine can fail if it is unable to load a collating sequence 1943 ** required for string comparison, or if unable to allocate memory 1944 ** for a UTF conversion required for comparison. The error is stored 1945 ** in the pParse structure. 1946 */ 1947 static int whereInScanEst( 1948 Parse *pParse, /* Parsing & code generating context */ 1949 WhereLoopBuilder *pBuilder, 1950 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1951 tRowcnt *pnRow /* Write the revised row estimate here */ 1952 ){ 1953 Index *p = pBuilder->pNew->u.btree.pIndex; 1954 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1955 int nRecValid = pBuilder->nRecValid; 1956 int rc = SQLITE_OK; /* Subfunction return code */ 1957 tRowcnt nEst; /* Number of rows for a single term */ 1958 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1959 int i; /* Loop counter */ 1960 1961 assert( p->aSample!=0 ); 1962 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1963 nEst = nRow0; 1964 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1965 nRowEst += nEst; 1966 pBuilder->nRecValid = nRecValid; 1967 } 1968 1969 if( rc==SQLITE_OK ){ 1970 if( nRowEst > nRow0 ) nRowEst = nRow0; 1971 *pnRow = nRowEst; 1972 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1973 } 1974 assert( pBuilder->nRecValid==nRecValid ); 1975 return rc; 1976 } 1977 #endif /* SQLITE_ENABLE_STAT4 */ 1978 1979 1980 #ifdef WHERETRACE_ENABLED 1981 /* 1982 ** Print the content of a WhereTerm object 1983 */ 1984 void sqlite3WhereTermPrint(WhereTerm *pTerm, int iTerm){ 1985 if( pTerm==0 ){ 1986 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1987 }else{ 1988 char zType[8]; 1989 char zLeft[50]; 1990 memcpy(zType, "....", 5); 1991 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1992 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1993 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1994 if( pTerm->wtFlags & TERM_CODED ) zType[3] = 'C'; 1995 if( pTerm->eOperator & WO_SINGLE ){ 1996 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 1997 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 1998 pTerm->leftCursor, pTerm->u.x.leftColumn); 1999 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 2000 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%llx", 2001 pTerm->u.pOrInfo->indexable); 2002 }else{ 2003 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 2004 } 2005 sqlite3DebugPrintf( 2006 "TERM-%-3d %p %s %-12s op=%03x wtFlags=%04x", 2007 iTerm, pTerm, zType, zLeft, pTerm->eOperator, pTerm->wtFlags); 2008 /* The 0x10000 .wheretrace flag causes extra information to be 2009 ** shown about each Term */ 2010 if( sqlite3WhereTrace & 0x10000 ){ 2011 sqlite3DebugPrintf(" prob=%-3d prereq=%llx,%llx", 2012 pTerm->truthProb, (u64)pTerm->prereqAll, (u64)pTerm->prereqRight); 2013 } 2014 if( (pTerm->eOperator & (WO_OR|WO_AND))==0 && pTerm->u.x.iField ){ 2015 sqlite3DebugPrintf(" iField=%d", pTerm->u.x.iField); 2016 } 2017 if( pTerm->iParent>=0 ){ 2018 sqlite3DebugPrintf(" iParent=%d", pTerm->iParent); 2019 } 2020 sqlite3DebugPrintf("\n"); 2021 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 2022 } 2023 } 2024 #endif 2025 2026 #ifdef WHERETRACE_ENABLED 2027 /* 2028 ** Show the complete content of a WhereClause 2029 */ 2030 void sqlite3WhereClausePrint(WhereClause *pWC){ 2031 int i; 2032 for(i=0; i<pWC->nTerm; i++){ 2033 sqlite3WhereTermPrint(&pWC->a[i], i); 2034 } 2035 } 2036 #endif 2037 2038 #ifdef WHERETRACE_ENABLED 2039 /* 2040 ** Print a WhereLoop object for debugging purposes 2041 */ 2042 void sqlite3WhereLoopPrint(WhereLoop *p, WhereClause *pWC){ 2043 WhereInfo *pWInfo = pWC->pWInfo; 2044 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 2045 SrcItem *pItem = pWInfo->pTabList->a + p->iTab; 2046 Table *pTab = pItem->pTab; 2047 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 2048 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 2049 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 2050 sqlite3DebugPrintf(" %12s", 2051 pItem->zAlias ? pItem->zAlias : pTab->zName); 2052 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2053 const char *zName; 2054 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 2055 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 2056 int i = sqlite3Strlen30(zName) - 1; 2057 while( zName[i]!='_' ) i--; 2058 zName += i; 2059 } 2060 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 2061 }else{ 2062 sqlite3DebugPrintf("%20s",""); 2063 } 2064 }else{ 2065 char *z; 2066 if( p->u.vtab.idxStr ){ 2067 z = sqlite3_mprintf("(%d,\"%s\",%#x)", 2068 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 2069 }else{ 2070 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 2071 } 2072 sqlite3DebugPrintf(" %-19s", z); 2073 sqlite3_free(z); 2074 } 2075 if( p->wsFlags & WHERE_SKIPSCAN ){ 2076 sqlite3DebugPrintf(" f %06x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 2077 }else{ 2078 sqlite3DebugPrintf(" f %06x N %d", p->wsFlags, p->nLTerm); 2079 } 2080 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 2081 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 2082 int i; 2083 for(i=0; i<p->nLTerm; i++){ 2084 sqlite3WhereTermPrint(p->aLTerm[i], i); 2085 } 2086 } 2087 } 2088 #endif 2089 2090 /* 2091 ** Convert bulk memory into a valid WhereLoop that can be passed 2092 ** to whereLoopClear harmlessly. 2093 */ 2094 static void whereLoopInit(WhereLoop *p){ 2095 p->aLTerm = p->aLTermSpace; 2096 p->nLTerm = 0; 2097 p->nLSlot = ArraySize(p->aLTermSpace); 2098 p->wsFlags = 0; 2099 } 2100 2101 /* 2102 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 2103 */ 2104 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 2105 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 2106 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 2107 sqlite3_free(p->u.vtab.idxStr); 2108 p->u.vtab.needFree = 0; 2109 p->u.vtab.idxStr = 0; 2110 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 2111 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 2112 sqlite3DbFreeNN(db, p->u.btree.pIndex); 2113 p->u.btree.pIndex = 0; 2114 } 2115 } 2116 } 2117 2118 /* 2119 ** Deallocate internal memory used by a WhereLoop object 2120 */ 2121 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 2122 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 2123 whereLoopClearUnion(db, p); 2124 whereLoopInit(p); 2125 } 2126 2127 /* 2128 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 2129 */ 2130 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 2131 WhereTerm **paNew; 2132 if( p->nLSlot>=n ) return SQLITE_OK; 2133 n = (n+7)&~7; 2134 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 2135 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 2136 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 2137 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 2138 p->aLTerm = paNew; 2139 p->nLSlot = n; 2140 return SQLITE_OK; 2141 } 2142 2143 /* 2144 ** Transfer content from the second pLoop into the first. 2145 */ 2146 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 2147 whereLoopClearUnion(db, pTo); 2148 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 2149 memset(pTo, 0, WHERE_LOOP_XFER_SZ); 2150 return SQLITE_NOMEM_BKPT; 2151 } 2152 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 2153 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 2154 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 2155 pFrom->u.vtab.needFree = 0; 2156 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 2157 pFrom->u.btree.pIndex = 0; 2158 } 2159 return SQLITE_OK; 2160 } 2161 2162 /* 2163 ** Delete a WhereLoop object 2164 */ 2165 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 2166 whereLoopClear(db, p); 2167 sqlite3DbFreeNN(db, p); 2168 } 2169 2170 /* 2171 ** Free a WhereInfo structure 2172 */ 2173 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 2174 int i; 2175 assert( pWInfo!=0 ); 2176 for(i=0; i<pWInfo->nLevel; i++){ 2177 WhereLevel *pLevel = &pWInfo->a[i]; 2178 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE)!=0 ){ 2179 assert( (pLevel->pWLoop->wsFlags & WHERE_MULTI_OR)==0 ); 2180 sqlite3DbFree(db, pLevel->u.in.aInLoop); 2181 } 2182 } 2183 sqlite3WhereClauseClear(&pWInfo->sWC); 2184 while( pWInfo->pLoops ){ 2185 WhereLoop *p = pWInfo->pLoops; 2186 pWInfo->pLoops = p->pNextLoop; 2187 whereLoopDelete(db, p); 2188 } 2189 assert( pWInfo->pExprMods==0 ); 2190 sqlite3DbFreeNN(db, pWInfo); 2191 } 2192 2193 /* Undo all Expr node modifications 2194 */ 2195 static void whereUndoExprMods(WhereInfo *pWInfo){ 2196 while( pWInfo->pExprMods ){ 2197 WhereExprMod *p = pWInfo->pExprMods; 2198 pWInfo->pExprMods = p->pNext; 2199 memcpy(p->pExpr, &p->orig, sizeof(p->orig)); 2200 sqlite3DbFree(pWInfo->pParse->db, p); 2201 } 2202 } 2203 2204 /* 2205 ** Return TRUE if all of the following are true: 2206 ** 2207 ** (1) X has the same or lower cost, or returns the same or fewer rows, 2208 ** than Y. 2209 ** (2) X uses fewer WHERE clause terms than Y 2210 ** (3) Every WHERE clause term used by X is also used by Y 2211 ** (4) X skips at least as many columns as Y 2212 ** (5) If X is a covering index, than Y is too 2213 ** 2214 ** Conditions (2) and (3) mean that X is a "proper subset" of Y. 2215 ** If X is a proper subset of Y then Y is a better choice and ought 2216 ** to have a lower cost. This routine returns TRUE when that cost 2217 ** relationship is inverted and needs to be adjusted. Constraint (4) 2218 ** was added because if X uses skip-scan less than Y it still might 2219 ** deserve a lower cost even if it is a proper subset of Y. Constraint (5) 2220 ** was added because a covering index probably deserves to have a lower cost 2221 ** than a non-covering index even if it is a proper subset. 2222 */ 2223 static int whereLoopCheaperProperSubset( 2224 const WhereLoop *pX, /* First WhereLoop to compare */ 2225 const WhereLoop *pY /* Compare against this WhereLoop */ 2226 ){ 2227 int i, j; 2228 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 2229 return 0; /* X is not a subset of Y */ 2230 } 2231 if( pX->rRun>pY->rRun && pX->nOut>pY->nOut ) return 0; 2232 if( pY->nSkip > pX->nSkip ) return 0; 2233 for(i=pX->nLTerm-1; i>=0; i--){ 2234 if( pX->aLTerm[i]==0 ) continue; 2235 for(j=pY->nLTerm-1; j>=0; j--){ 2236 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 2237 } 2238 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 2239 } 2240 if( (pX->wsFlags&WHERE_IDX_ONLY)!=0 2241 && (pY->wsFlags&WHERE_IDX_ONLY)==0 ){ 2242 return 0; /* Constraint (5) */ 2243 } 2244 return 1; /* All conditions meet */ 2245 } 2246 2247 /* 2248 ** Try to adjust the cost and number of output rows of WhereLoop pTemplate 2249 ** upwards or downwards so that: 2250 ** 2251 ** (1) pTemplate costs less than any other WhereLoops that are a proper 2252 ** subset of pTemplate 2253 ** 2254 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 2255 ** is a proper subset. 2256 ** 2257 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 2258 ** WHERE clause terms than Y and that every WHERE clause term used by X is 2259 ** also used by Y. 2260 */ 2261 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 2262 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 2263 for(; p; p=p->pNextLoop){ 2264 if( p->iTab!=pTemplate->iTab ) continue; 2265 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 2266 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 2267 /* Adjust pTemplate cost downward so that it is cheaper than its 2268 ** subset p. */ 2269 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2270 pTemplate->rRun, pTemplate->nOut, 2271 MIN(p->rRun, pTemplate->rRun), 2272 MIN(p->nOut - 1, pTemplate->nOut))); 2273 pTemplate->rRun = MIN(p->rRun, pTemplate->rRun); 2274 pTemplate->nOut = MIN(p->nOut - 1, pTemplate->nOut); 2275 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 2276 /* Adjust pTemplate cost upward so that it is costlier than p since 2277 ** pTemplate is a proper subset of p */ 2278 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2279 pTemplate->rRun, pTemplate->nOut, 2280 MAX(p->rRun, pTemplate->rRun), 2281 MAX(p->nOut + 1, pTemplate->nOut))); 2282 pTemplate->rRun = MAX(p->rRun, pTemplate->rRun); 2283 pTemplate->nOut = MAX(p->nOut + 1, pTemplate->nOut); 2284 } 2285 } 2286 } 2287 2288 /* 2289 ** Search the list of WhereLoops in *ppPrev looking for one that can be 2290 ** replaced by pTemplate. 2291 ** 2292 ** Return NULL if pTemplate does not belong on the WhereLoop list. 2293 ** In other words if pTemplate ought to be dropped from further consideration. 2294 ** 2295 ** If pX is a WhereLoop that pTemplate can replace, then return the 2296 ** link that points to pX. 2297 ** 2298 ** If pTemplate cannot replace any existing element of the list but needs 2299 ** to be added to the list as a new entry, then return a pointer to the 2300 ** tail of the list. 2301 */ 2302 static WhereLoop **whereLoopFindLesser( 2303 WhereLoop **ppPrev, 2304 const WhereLoop *pTemplate 2305 ){ 2306 WhereLoop *p; 2307 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 2308 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 2309 /* If either the iTab or iSortIdx values for two WhereLoop are different 2310 ** then those WhereLoops need to be considered separately. Neither is 2311 ** a candidate to replace the other. */ 2312 continue; 2313 } 2314 /* In the current implementation, the rSetup value is either zero 2315 ** or the cost of building an automatic index (NlogN) and the NlogN 2316 ** is the same for compatible WhereLoops. */ 2317 assert( p->rSetup==0 || pTemplate->rSetup==0 2318 || p->rSetup==pTemplate->rSetup ); 2319 2320 /* whereLoopAddBtree() always generates and inserts the automatic index 2321 ** case first. Hence compatible candidate WhereLoops never have a larger 2322 ** rSetup. Call this SETUP-INVARIANT */ 2323 assert( p->rSetup>=pTemplate->rSetup ); 2324 2325 /* Any loop using an appliation-defined index (or PRIMARY KEY or 2326 ** UNIQUE constraint) with one or more == constraints is better 2327 ** than an automatic index. Unless it is a skip-scan. */ 2328 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 2329 && (pTemplate->nSkip)==0 2330 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 2331 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 2332 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2333 ){ 2334 break; 2335 } 2336 2337 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2338 ** discarded. WhereLoop p is better if: 2339 ** (1) p has no more dependencies than pTemplate, and 2340 ** (2) p has an equal or lower cost than pTemplate 2341 */ 2342 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2343 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2344 && p->rRun<=pTemplate->rRun /* (2b) */ 2345 && p->nOut<=pTemplate->nOut /* (2c) */ 2346 ){ 2347 return 0; /* Discard pTemplate */ 2348 } 2349 2350 /* If pTemplate is always better than p, then cause p to be overwritten 2351 ** with pTemplate. pTemplate is better than p if: 2352 ** (1) pTemplate has no more dependences than p, and 2353 ** (2) pTemplate has an equal or lower cost than p. 2354 */ 2355 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2356 && p->rRun>=pTemplate->rRun /* (2a) */ 2357 && p->nOut>=pTemplate->nOut /* (2b) */ 2358 ){ 2359 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2360 break; /* Cause p to be overwritten by pTemplate */ 2361 } 2362 } 2363 return ppPrev; 2364 } 2365 2366 /* 2367 ** Insert or replace a WhereLoop entry using the template supplied. 2368 ** 2369 ** An existing WhereLoop entry might be overwritten if the new template 2370 ** is better and has fewer dependencies. Or the template will be ignored 2371 ** and no insert will occur if an existing WhereLoop is faster and has 2372 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2373 ** added based on the template. 2374 ** 2375 ** If pBuilder->pOrSet is not NULL then we care about only the 2376 ** prerequisites and rRun and nOut costs of the N best loops. That 2377 ** information is gathered in the pBuilder->pOrSet object. This special 2378 ** processing mode is used only for OR clause processing. 2379 ** 2380 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2381 ** still might overwrite similar loops with the new template if the 2382 ** new template is better. Loops may be overwritten if the following 2383 ** conditions are met: 2384 ** 2385 ** (1) They have the same iTab. 2386 ** (2) They have the same iSortIdx. 2387 ** (3) The template has same or fewer dependencies than the current loop 2388 ** (4) The template has the same or lower cost than the current loop 2389 */ 2390 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2391 WhereLoop **ppPrev, *p; 2392 WhereInfo *pWInfo = pBuilder->pWInfo; 2393 sqlite3 *db = pWInfo->pParse->db; 2394 int rc; 2395 2396 /* Stop the search once we hit the query planner search limit */ 2397 if( pBuilder->iPlanLimit==0 ){ 2398 WHERETRACE(0xffffffff,("=== query planner search limit reached ===\n")); 2399 if( pBuilder->pOrSet ) pBuilder->pOrSet->n = 0; 2400 return SQLITE_DONE; 2401 } 2402 pBuilder->iPlanLimit--; 2403 2404 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2405 2406 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2407 ** and prereqs. 2408 */ 2409 if( pBuilder->pOrSet!=0 ){ 2410 if( pTemplate->nLTerm ){ 2411 #if WHERETRACE_ENABLED 2412 u16 n = pBuilder->pOrSet->n; 2413 int x = 2414 #endif 2415 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2416 pTemplate->nOut); 2417 #if WHERETRACE_ENABLED /* 0x8 */ 2418 if( sqlite3WhereTrace & 0x8 ){ 2419 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2420 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2421 } 2422 #endif 2423 } 2424 return SQLITE_OK; 2425 } 2426 2427 /* Look for an existing WhereLoop to replace with pTemplate 2428 */ 2429 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2430 2431 if( ppPrev==0 ){ 2432 /* There already exists a WhereLoop on the list that is better 2433 ** than pTemplate, so just ignore pTemplate */ 2434 #if WHERETRACE_ENABLED /* 0x8 */ 2435 if( sqlite3WhereTrace & 0x8 ){ 2436 sqlite3DebugPrintf(" skip: "); 2437 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2438 } 2439 #endif 2440 return SQLITE_OK; 2441 }else{ 2442 p = *ppPrev; 2443 } 2444 2445 /* If we reach this point it means that either p[] should be overwritten 2446 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2447 ** WhereLoop and insert it. 2448 */ 2449 #if WHERETRACE_ENABLED /* 0x8 */ 2450 if( sqlite3WhereTrace & 0x8 ){ 2451 if( p!=0 ){ 2452 sqlite3DebugPrintf("replace: "); 2453 sqlite3WhereLoopPrint(p, pBuilder->pWC); 2454 sqlite3DebugPrintf(" with: "); 2455 }else{ 2456 sqlite3DebugPrintf(" add: "); 2457 } 2458 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2459 } 2460 #endif 2461 if( p==0 ){ 2462 /* Allocate a new WhereLoop to add to the end of the list */ 2463 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2464 if( p==0 ) return SQLITE_NOMEM_BKPT; 2465 whereLoopInit(p); 2466 p->pNextLoop = 0; 2467 }else{ 2468 /* We will be overwriting WhereLoop p[]. But before we do, first 2469 ** go through the rest of the list and delete any other entries besides 2470 ** p[] that are also supplated by pTemplate */ 2471 WhereLoop **ppTail = &p->pNextLoop; 2472 WhereLoop *pToDel; 2473 while( *ppTail ){ 2474 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2475 if( ppTail==0 ) break; 2476 pToDel = *ppTail; 2477 if( pToDel==0 ) break; 2478 *ppTail = pToDel->pNextLoop; 2479 #if WHERETRACE_ENABLED /* 0x8 */ 2480 if( sqlite3WhereTrace & 0x8 ){ 2481 sqlite3DebugPrintf(" delete: "); 2482 sqlite3WhereLoopPrint(pToDel, pBuilder->pWC); 2483 } 2484 #endif 2485 whereLoopDelete(db, pToDel); 2486 } 2487 } 2488 rc = whereLoopXfer(db, p, pTemplate); 2489 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2490 Index *pIndex = p->u.btree.pIndex; 2491 if( pIndex && pIndex->idxType==SQLITE_IDXTYPE_IPK ){ 2492 p->u.btree.pIndex = 0; 2493 } 2494 } 2495 return rc; 2496 } 2497 2498 /* 2499 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2500 ** WHERE clause that reference the loop but which are not used by an 2501 ** index. 2502 * 2503 ** For every WHERE clause term that is not used by the index 2504 ** and which has a truth probability assigned by one of the likelihood(), 2505 ** likely(), or unlikely() SQL functions, reduce the estimated number 2506 ** of output rows by the probability specified. 2507 ** 2508 ** TUNING: For every WHERE clause term that is not used by the index 2509 ** and which does not have an assigned truth probability, heuristics 2510 ** described below are used to try to estimate the truth probability. 2511 ** TODO --> Perhaps this is something that could be improved by better 2512 ** table statistics. 2513 ** 2514 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2515 ** value corresponds to -1 in LogEst notation, so this means decrement 2516 ** the WhereLoop.nOut field for every such WHERE clause term. 2517 ** 2518 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2519 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2520 ** final output row estimate is no greater than 1/4 of the total number 2521 ** of rows in the table. In other words, assume that x==EXPR will filter 2522 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2523 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2524 ** on the "x" column and so in that case only cap the output row estimate 2525 ** at 1/2 instead of 1/4. 2526 */ 2527 static void whereLoopOutputAdjust( 2528 WhereClause *pWC, /* The WHERE clause */ 2529 WhereLoop *pLoop, /* The loop to adjust downward */ 2530 LogEst nRow /* Number of rows in the entire table */ 2531 ){ 2532 WhereTerm *pTerm, *pX; 2533 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2534 int i, j; 2535 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2536 2537 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2538 for(i=pWC->nBase, pTerm=pWC->a; i>0; i--, pTerm++){ 2539 assert( pTerm!=0 ); 2540 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2541 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2542 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) continue; 2543 for(j=pLoop->nLTerm-1; j>=0; j--){ 2544 pX = pLoop->aLTerm[j]; 2545 if( pX==0 ) continue; 2546 if( pX==pTerm ) break; 2547 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2548 } 2549 if( j<0 ){ 2550 if( pLoop->maskSelf==pTerm->prereqAll ){ 2551 /* If there are extra terms in the WHERE clause not used by an index 2552 ** that depend only on the table being scanned, and that will tend to 2553 ** cause many rows to be omitted, then mark that table as 2554 ** "self-culling". */ 2555 pLoop->wsFlags |= WHERE_SELFCULL; 2556 } 2557 if( pTerm->truthProb<=0 ){ 2558 /* If a truth probability is specified using the likelihood() hints, 2559 ** then use the probability provided by the application. */ 2560 pLoop->nOut += pTerm->truthProb; 2561 }else{ 2562 /* In the absence of explicit truth probabilities, use heuristics to 2563 ** guess a reasonable truth probability. */ 2564 pLoop->nOut--; 2565 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 2566 && (pTerm->wtFlags & TERM_HIGHTRUTH)==0 /* tag-20200224-1 */ 2567 ){ 2568 Expr *pRight = pTerm->pExpr->pRight; 2569 int k = 0; 2570 testcase( pTerm->pExpr->op==TK_IS ); 2571 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2572 k = 10; 2573 }else{ 2574 k = 20; 2575 } 2576 if( iReduce<k ){ 2577 pTerm->wtFlags |= TERM_HEURTRUTH; 2578 iReduce = k; 2579 } 2580 } 2581 } 2582 } 2583 } 2584 if( pLoop->nOut > nRow-iReduce ){ 2585 pLoop->nOut = nRow - iReduce; 2586 } 2587 } 2588 2589 /* 2590 ** Term pTerm is a vector range comparison operation. The first comparison 2591 ** in the vector can be optimized using column nEq of the index. This 2592 ** function returns the total number of vector elements that can be used 2593 ** as part of the range comparison. 2594 ** 2595 ** For example, if the query is: 2596 ** 2597 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2598 ** 2599 ** and the index: 2600 ** 2601 ** CREATE INDEX ... ON (a, b, c, d, e) 2602 ** 2603 ** then this function would be invoked with nEq=1. The value returned in 2604 ** this case is 3. 2605 */ 2606 static int whereRangeVectorLen( 2607 Parse *pParse, /* Parsing context */ 2608 int iCur, /* Cursor open on pIdx */ 2609 Index *pIdx, /* The index to be used for a inequality constraint */ 2610 int nEq, /* Number of prior equality constraints on same index */ 2611 WhereTerm *pTerm /* The vector inequality constraint */ 2612 ){ 2613 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2614 int i; 2615 2616 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2617 for(i=1; i<nCmp; i++){ 2618 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2619 ** of the index. If not, exit the loop. */ 2620 char aff; /* Comparison affinity */ 2621 char idxaff = 0; /* Indexed columns affinity */ 2622 CollSeq *pColl; /* Comparison collation sequence */ 2623 Expr *pLhs, *pRhs; 2624 2625 assert( ExprUseXList(pTerm->pExpr->pLeft) ); 2626 pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2627 pRhs = pTerm->pExpr->pRight; 2628 if( ExprUseXSelect(pRhs) ){ 2629 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2630 }else{ 2631 pRhs = pRhs->x.pList->a[i].pExpr; 2632 } 2633 2634 /* Check that the LHS of the comparison is a column reference to 2635 ** the right column of the right source table. And that the sort 2636 ** order of the index column is the same as the sort order of the 2637 ** leftmost index column. */ 2638 if( pLhs->op!=TK_COLUMN 2639 || pLhs->iTable!=iCur 2640 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2641 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2642 ){ 2643 break; 2644 } 2645 2646 testcase( pLhs->iColumn==XN_ROWID ); 2647 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2648 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2649 if( aff!=idxaff ) break; 2650 2651 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2652 if( pColl==0 ) break; 2653 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2654 } 2655 return i; 2656 } 2657 2658 /* 2659 ** Adjust the cost C by the costMult facter T. This only occurs if 2660 ** compiled with -DSQLITE_ENABLE_COSTMULT 2661 */ 2662 #ifdef SQLITE_ENABLE_COSTMULT 2663 # define ApplyCostMultiplier(C,T) C += T 2664 #else 2665 # define ApplyCostMultiplier(C,T) 2666 #endif 2667 2668 /* 2669 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2670 ** index pIndex. Try to match one more. 2671 ** 2672 ** When this function is called, pBuilder->pNew->nOut contains the 2673 ** number of rows expected to be visited by filtering using the nEq 2674 ** terms only. If it is modified, this value is restored before this 2675 ** function returns. 2676 ** 2677 ** If pProbe->idxType==SQLITE_IDXTYPE_IPK, that means pIndex is 2678 ** a fake index used for the INTEGER PRIMARY KEY. 2679 */ 2680 static int whereLoopAddBtreeIndex( 2681 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2682 SrcItem *pSrc, /* FROM clause term being analyzed */ 2683 Index *pProbe, /* An index on pSrc */ 2684 LogEst nInMul /* log(Number of iterations due to IN) */ 2685 ){ 2686 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2687 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2688 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2689 WhereLoop *pNew; /* Template WhereLoop under construction */ 2690 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2691 int opMask; /* Valid operators for constraints */ 2692 WhereScan scan; /* Iterator for WHERE terms */ 2693 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2694 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2695 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2696 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2697 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2698 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2699 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2700 LogEst saved_nOut; /* Original value of pNew->nOut */ 2701 int rc = SQLITE_OK; /* Return code */ 2702 LogEst rSize; /* Number of rows in the table */ 2703 LogEst rLogSize; /* Logarithm of table size */ 2704 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2705 2706 pNew = pBuilder->pNew; 2707 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2708 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d, nSkip=%d, rRun=%d\n", 2709 pProbe->pTable->zName,pProbe->zName, 2710 pNew->u.btree.nEq, pNew->nSkip, pNew->rRun)); 2711 2712 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2713 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2714 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2715 opMask = WO_LT|WO_LE; 2716 }else{ 2717 assert( pNew->u.btree.nBtm==0 ); 2718 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2719 } 2720 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2721 2722 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2723 assert( pNew->u.btree.nEq<pProbe->nKeyCol 2724 || pProbe->idxType!=SQLITE_IDXTYPE_PRIMARYKEY ); 2725 2726 saved_nEq = pNew->u.btree.nEq; 2727 saved_nBtm = pNew->u.btree.nBtm; 2728 saved_nTop = pNew->u.btree.nTop; 2729 saved_nSkip = pNew->nSkip; 2730 saved_nLTerm = pNew->nLTerm; 2731 saved_wsFlags = pNew->wsFlags; 2732 saved_prereq = pNew->prereq; 2733 saved_nOut = pNew->nOut; 2734 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2735 opMask, pProbe); 2736 pNew->rSetup = 0; 2737 rSize = pProbe->aiRowLogEst[0]; 2738 rLogSize = estLog(rSize); 2739 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2740 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2741 LogEst rCostIdx; 2742 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2743 int nIn = 0; 2744 #ifdef SQLITE_ENABLE_STAT4 2745 int nRecValid = pBuilder->nRecValid; 2746 #endif 2747 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2748 && indexColumnNotNull(pProbe, saved_nEq) 2749 ){ 2750 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2751 } 2752 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2753 2754 /* Do not allow the upper bound of a LIKE optimization range constraint 2755 ** to mix with a lower range bound from some other source */ 2756 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2757 2758 /* tag-20191211-001: Do not allow constraints from the WHERE clause to 2759 ** be used by the right table of a LEFT JOIN. Only constraints in the 2760 ** ON clause are allowed. See tag-20191211-002 for the vtab equivalent. */ 2761 if( (pSrc->fg.jointype & JT_LEFT)!=0 2762 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2763 ){ 2764 continue; 2765 } 2766 2767 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2768 pBuilder->bldFlags1 |= SQLITE_BLDF1_UNIQUE; 2769 }else{ 2770 pBuilder->bldFlags1 |= SQLITE_BLDF1_INDEXED; 2771 } 2772 pNew->wsFlags = saved_wsFlags; 2773 pNew->u.btree.nEq = saved_nEq; 2774 pNew->u.btree.nBtm = saved_nBtm; 2775 pNew->u.btree.nTop = saved_nTop; 2776 pNew->nLTerm = saved_nLTerm; 2777 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2778 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2779 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2780 2781 assert( nInMul==0 2782 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2783 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2784 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2785 ); 2786 2787 if( eOp & WO_IN ){ 2788 Expr *pExpr = pTerm->pExpr; 2789 if( ExprUseXSelect(pExpr) ){ 2790 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2791 int i; 2792 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2793 2794 /* The expression may actually be of the form (x, y) IN (SELECT...). 2795 ** In this case there is a separate term for each of (x) and (y). 2796 ** However, the nIn multiplier should only be applied once, not once 2797 ** for each such term. The following loop checks that pTerm is the 2798 ** first such term in use, and sets nIn back to 0 if it is not. */ 2799 for(i=0; i<pNew->nLTerm-1; i++){ 2800 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2801 } 2802 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2803 /* "x IN (value, value, ...)" */ 2804 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2805 } 2806 if( pProbe->hasStat1 && rLogSize>=10 ){ 2807 LogEst M, logK, x; 2808 /* Let: 2809 ** N = the total number of rows in the table 2810 ** K = the number of entries on the RHS of the IN operator 2811 ** M = the number of rows in the table that match terms to the 2812 ** to the left in the same index. If the IN operator is on 2813 ** the left-most index column, M==N. 2814 ** 2815 ** Given the definitions above, it is better to omit the IN operator 2816 ** from the index lookup and instead do a scan of the M elements, 2817 ** testing each scanned row against the IN operator separately, if: 2818 ** 2819 ** M*log(K) < K*log(N) 2820 ** 2821 ** Our estimates for M, K, and N might be inaccurate, so we build in 2822 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator 2823 ** with the index, as using an index has better worst-case behavior. 2824 ** If we do not have real sqlite_stat1 data, always prefer to use 2825 ** the index. Do not bother with this optimization on very small 2826 ** tables (less than 2 rows) as it is pointless in that case. 2827 */ 2828 M = pProbe->aiRowLogEst[saved_nEq]; 2829 logK = estLog(nIn); 2830 /* TUNING v----- 10 to bias toward indexed IN */ 2831 x = M + logK + 10 - (nIn + rLogSize); 2832 if( x>=0 ){ 2833 WHERETRACE(0x40, 2834 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d) " 2835 "prefers indexed lookup\n", 2836 saved_nEq, M, logK, nIn, rLogSize, x)); 2837 }else if( nInMul<2 && OptimizationEnabled(db, SQLITE_SeekScan) ){ 2838 WHERETRACE(0x40, 2839 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d" 2840 " nInMul=%d) prefers skip-scan\n", 2841 saved_nEq, M, logK, nIn, rLogSize, x, nInMul)); 2842 pNew->wsFlags |= WHERE_IN_SEEKSCAN; 2843 }else{ 2844 WHERETRACE(0x40, 2845 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d" 2846 " nInMul=%d) prefers normal scan\n", 2847 saved_nEq, M, logK, nIn, rLogSize, x, nInMul)); 2848 continue; 2849 } 2850 } 2851 pNew->wsFlags |= WHERE_COLUMN_IN; 2852 }else if( eOp & (WO_EQ|WO_IS) ){ 2853 int iCol = pProbe->aiColumn[saved_nEq]; 2854 pNew->wsFlags |= WHERE_COLUMN_EQ; 2855 assert( saved_nEq==pNew->u.btree.nEq ); 2856 if( iCol==XN_ROWID 2857 || (iCol>=0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2858 ){ 2859 if( iCol==XN_ROWID || pProbe->uniqNotNull 2860 || (pProbe->nKeyCol==1 && pProbe->onError && eOp==WO_EQ) 2861 ){ 2862 pNew->wsFlags |= WHERE_ONEROW; 2863 }else{ 2864 pNew->wsFlags |= WHERE_UNQ_WANTED; 2865 } 2866 } 2867 if( scan.iEquiv>1 ) pNew->wsFlags |= WHERE_TRANSCONS; 2868 }else if( eOp & WO_ISNULL ){ 2869 pNew->wsFlags |= WHERE_COLUMN_NULL; 2870 }else if( eOp & (WO_GT|WO_GE) ){ 2871 testcase( eOp & WO_GT ); 2872 testcase( eOp & WO_GE ); 2873 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2874 pNew->u.btree.nBtm = whereRangeVectorLen( 2875 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2876 ); 2877 pBtm = pTerm; 2878 pTop = 0; 2879 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2880 /* Range constraints that come from the LIKE optimization are 2881 ** always used in pairs. */ 2882 pTop = &pTerm[1]; 2883 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2884 assert( pTop->wtFlags & TERM_LIKEOPT ); 2885 assert( pTop->eOperator==WO_LT ); 2886 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2887 pNew->aLTerm[pNew->nLTerm++] = pTop; 2888 pNew->wsFlags |= WHERE_TOP_LIMIT; 2889 pNew->u.btree.nTop = 1; 2890 } 2891 }else{ 2892 assert( eOp & (WO_LT|WO_LE) ); 2893 testcase( eOp & WO_LT ); 2894 testcase( eOp & WO_LE ); 2895 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2896 pNew->u.btree.nTop = whereRangeVectorLen( 2897 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2898 ); 2899 pTop = pTerm; 2900 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2901 pNew->aLTerm[pNew->nLTerm-2] : 0; 2902 } 2903 2904 /* At this point pNew->nOut is set to the number of rows expected to 2905 ** be visited by the index scan before considering term pTerm, or the 2906 ** values of nIn and nInMul. In other words, assuming that all 2907 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2908 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2909 assert( pNew->nOut==saved_nOut ); 2910 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2911 /* Adjust nOut using stat4 data. Or, if there is no stat4 2912 ** data, using some other estimate. */ 2913 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2914 }else{ 2915 int nEq = ++pNew->u.btree.nEq; 2916 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2917 2918 assert( pNew->nOut==saved_nOut ); 2919 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2920 assert( (eOp & WO_IN) || nIn==0 ); 2921 testcase( eOp & WO_IN ); 2922 pNew->nOut += pTerm->truthProb; 2923 pNew->nOut -= nIn; 2924 }else{ 2925 #ifdef SQLITE_ENABLE_STAT4 2926 tRowcnt nOut = 0; 2927 if( nInMul==0 2928 && pProbe->nSample 2929 && ALWAYS(pNew->u.btree.nEq<=pProbe->nSampleCol) 2930 && ((eOp & WO_IN)==0 || ExprUseXList(pTerm->pExpr)) 2931 && OptimizationEnabled(db, SQLITE_Stat4) 2932 ){ 2933 Expr *pExpr = pTerm->pExpr; 2934 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2935 testcase( eOp & WO_EQ ); 2936 testcase( eOp & WO_IS ); 2937 testcase( eOp & WO_ISNULL ); 2938 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2939 }else{ 2940 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2941 } 2942 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2943 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2944 if( nOut ){ 2945 pNew->nOut = sqlite3LogEst(nOut); 2946 if( nEq==1 2947 /* TUNING: Mark terms as "low selectivity" if they seem likely 2948 ** to be true for half or more of the rows in the table. 2949 ** See tag-202002240-1 */ 2950 && pNew->nOut+10 > pProbe->aiRowLogEst[0] 2951 ){ 2952 #if WHERETRACE_ENABLED /* 0x01 */ 2953 if( sqlite3WhereTrace & 0x01 ){ 2954 sqlite3DebugPrintf( 2955 "STAT4 determines term has low selectivity:\n"); 2956 sqlite3WhereTermPrint(pTerm, 999); 2957 } 2958 #endif 2959 pTerm->wtFlags |= TERM_HIGHTRUTH; 2960 if( pTerm->wtFlags & TERM_HEURTRUTH ){ 2961 /* If the term has previously been used with an assumption of 2962 ** higher selectivity, then set the flag to rerun the 2963 ** loop computations. */ 2964 pBuilder->bldFlags2 |= SQLITE_BLDF2_2NDPASS; 2965 } 2966 } 2967 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2968 pNew->nOut -= nIn; 2969 } 2970 } 2971 if( nOut==0 ) 2972 #endif 2973 { 2974 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2975 if( eOp & WO_ISNULL ){ 2976 /* TUNING: If there is no likelihood() value, assume that a 2977 ** "col IS NULL" expression matches twice as many rows 2978 ** as (col=?). */ 2979 pNew->nOut += 10; 2980 } 2981 } 2982 } 2983 } 2984 2985 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2986 ** it to pNew->rRun, which is currently set to the cost of the index 2987 ** seek only. Then, if this is a non-covering index, add the cost of 2988 ** visiting the rows in the main table. */ 2989 assert( pSrc->pTab->szTabRow>0 ); 2990 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2991 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2992 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2993 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2994 } 2995 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2996 2997 nOutUnadjusted = pNew->nOut; 2998 pNew->rRun += nInMul + nIn; 2999 pNew->nOut += nInMul + nIn; 3000 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 3001 rc = whereLoopInsert(pBuilder, pNew); 3002 3003 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 3004 pNew->nOut = saved_nOut; 3005 }else{ 3006 pNew->nOut = nOutUnadjusted; 3007 } 3008 3009 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 3010 && pNew->u.btree.nEq<pProbe->nColumn 3011 && (pNew->u.btree.nEq<pProbe->nKeyCol || 3012 pProbe->idxType!=SQLITE_IDXTYPE_PRIMARYKEY) 3013 ){ 3014 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 3015 } 3016 pNew->nOut = saved_nOut; 3017 #ifdef SQLITE_ENABLE_STAT4 3018 pBuilder->nRecValid = nRecValid; 3019 #endif 3020 } 3021 pNew->prereq = saved_prereq; 3022 pNew->u.btree.nEq = saved_nEq; 3023 pNew->u.btree.nBtm = saved_nBtm; 3024 pNew->u.btree.nTop = saved_nTop; 3025 pNew->nSkip = saved_nSkip; 3026 pNew->wsFlags = saved_wsFlags; 3027 pNew->nOut = saved_nOut; 3028 pNew->nLTerm = saved_nLTerm; 3029 3030 /* Consider using a skip-scan if there are no WHERE clause constraints 3031 ** available for the left-most terms of the index, and if the average 3032 ** number of repeats in the left-most terms is at least 18. 3033 ** 3034 ** The magic number 18 is selected on the basis that scanning 17 rows 3035 ** is almost always quicker than an index seek (even though if the index 3036 ** contains fewer than 2^17 rows we assume otherwise in other parts of 3037 ** the code). And, even if it is not, it should not be too much slower. 3038 ** On the other hand, the extra seeks could end up being significantly 3039 ** more expensive. */ 3040 assert( 42==sqlite3LogEst(18) ); 3041 if( saved_nEq==saved_nSkip 3042 && saved_nEq+1<pProbe->nKeyCol 3043 && saved_nEq==pNew->nLTerm 3044 && pProbe->noSkipScan==0 3045 && pProbe->hasStat1!=0 3046 && OptimizationEnabled(db, SQLITE_SkipScan) 3047 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 3048 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 3049 ){ 3050 LogEst nIter; 3051 pNew->u.btree.nEq++; 3052 pNew->nSkip++; 3053 pNew->aLTerm[pNew->nLTerm++] = 0; 3054 pNew->wsFlags |= WHERE_SKIPSCAN; 3055 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 3056 pNew->nOut -= nIter; 3057 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 3058 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 3059 nIter += 5; 3060 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 3061 pNew->nOut = saved_nOut; 3062 pNew->u.btree.nEq = saved_nEq; 3063 pNew->nSkip = saved_nSkip; 3064 pNew->wsFlags = saved_wsFlags; 3065 } 3066 3067 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n", 3068 pProbe->pTable->zName, pProbe->zName, saved_nEq, rc)); 3069 return rc; 3070 } 3071 3072 /* 3073 ** Return True if it is possible that pIndex might be useful in 3074 ** implementing the ORDER BY clause in pBuilder. 3075 ** 3076 ** Return False if pBuilder does not contain an ORDER BY clause or 3077 ** if there is no way for pIndex to be useful in implementing that 3078 ** ORDER BY clause. 3079 */ 3080 static int indexMightHelpWithOrderBy( 3081 WhereLoopBuilder *pBuilder, 3082 Index *pIndex, 3083 int iCursor 3084 ){ 3085 ExprList *pOB; 3086 ExprList *aColExpr; 3087 int ii, jj; 3088 3089 if( pIndex->bUnordered ) return 0; 3090 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 3091 for(ii=0; ii<pOB->nExpr; ii++){ 3092 Expr *pExpr = sqlite3ExprSkipCollateAndLikely(pOB->a[ii].pExpr); 3093 if( NEVER(pExpr==0) ) continue; 3094 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 3095 if( pExpr->iColumn<0 ) return 1; 3096 for(jj=0; jj<pIndex->nKeyCol; jj++){ 3097 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 3098 } 3099 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 3100 for(jj=0; jj<pIndex->nKeyCol; jj++){ 3101 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 3102 if( sqlite3ExprCompareSkip(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 3103 return 1; 3104 } 3105 } 3106 } 3107 } 3108 return 0; 3109 } 3110 3111 /* Check to see if a partial index with pPartIndexWhere can be used 3112 ** in the current query. Return true if it can be and false if not. 3113 */ 3114 static int whereUsablePartialIndex( 3115 int iTab, /* The table for which we want an index */ 3116 int isLeft, /* True if iTab is the right table of a LEFT JOIN */ 3117 WhereClause *pWC, /* The WHERE clause of the query */ 3118 Expr *pWhere /* The WHERE clause from the partial index */ 3119 ){ 3120 int i; 3121 WhereTerm *pTerm; 3122 Parse *pParse = pWC->pWInfo->pParse; 3123 while( pWhere->op==TK_AND ){ 3124 if( !whereUsablePartialIndex(iTab,isLeft,pWC,pWhere->pLeft) ) return 0; 3125 pWhere = pWhere->pRight; 3126 } 3127 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 3128 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 3129 Expr *pExpr; 3130 pExpr = pTerm->pExpr; 3131 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 3132 && (isLeft==0 || ExprHasProperty(pExpr, EP_FromJoin)) 3133 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 3134 && (pTerm->wtFlags & TERM_VNULL)==0 3135 ){ 3136 return 1; 3137 } 3138 } 3139 return 0; 3140 } 3141 3142 /* 3143 ** Add all WhereLoop objects for a single table of the join where the table 3144 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 3145 ** a b-tree table, not a virtual table. 3146 ** 3147 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 3148 ** are calculated as follows: 3149 ** 3150 ** For a full scan, assuming the table (or index) contains nRow rows: 3151 ** 3152 ** cost = nRow * 3.0 // full-table scan 3153 ** cost = nRow * K // scan of covering index 3154 ** cost = nRow * (K+3.0) // scan of non-covering index 3155 ** 3156 ** where K is a value between 1.1 and 3.0 set based on the relative 3157 ** estimated average size of the index and table records. 3158 ** 3159 ** For an index scan, where nVisit is the number of index rows visited 3160 ** by the scan, and nSeek is the number of seek operations required on 3161 ** the index b-tree: 3162 ** 3163 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 3164 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 3165 ** 3166 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 3167 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 3168 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 3169 ** 3170 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 3171 ** of uncertainty. For this reason, scoring is designed to pick plans that 3172 ** "do the least harm" if the estimates are inaccurate. For example, a 3173 ** log(nRow) factor is omitted from a non-covering index scan in order to 3174 ** bias the scoring in favor of using an index, since the worst-case 3175 ** performance of using an index is far better than the worst-case performance 3176 ** of a full table scan. 3177 */ 3178 static int whereLoopAddBtree( 3179 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3180 Bitmask mPrereq /* Extra prerequesites for using this table */ 3181 ){ 3182 WhereInfo *pWInfo; /* WHERE analysis context */ 3183 Index *pProbe; /* An index we are evaluating */ 3184 Index sPk; /* A fake index object for the primary key */ 3185 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 3186 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 3187 SrcList *pTabList; /* The FROM clause */ 3188 SrcItem *pSrc; /* The FROM clause btree term to add */ 3189 WhereLoop *pNew; /* Template WhereLoop object */ 3190 int rc = SQLITE_OK; /* Return code */ 3191 int iSortIdx = 1; /* Index number */ 3192 int b; /* A boolean value */ 3193 LogEst rSize; /* number of rows in the table */ 3194 WhereClause *pWC; /* The parsed WHERE clause */ 3195 Table *pTab; /* Table being queried */ 3196 3197 pNew = pBuilder->pNew; 3198 pWInfo = pBuilder->pWInfo; 3199 pTabList = pWInfo->pTabList; 3200 pSrc = pTabList->a + pNew->iTab; 3201 pTab = pSrc->pTab; 3202 pWC = pBuilder->pWC; 3203 assert( !IsVirtual(pSrc->pTab) ); 3204 3205 if( pSrc->fg.isIndexedBy ){ 3206 assert( pSrc->fg.isCte==0 ); 3207 /* An INDEXED BY clause specifies a particular index to use */ 3208 pProbe = pSrc->u2.pIBIndex; 3209 }else if( !HasRowid(pTab) ){ 3210 pProbe = pTab->pIndex; 3211 }else{ 3212 /* There is no INDEXED BY clause. Create a fake Index object in local 3213 ** variable sPk to represent the rowid primary key index. Make this 3214 ** fake index the first in a chain of Index objects with all of the real 3215 ** indices to follow */ 3216 Index *pFirst; /* First of real indices on the table */ 3217 memset(&sPk, 0, sizeof(Index)); 3218 sPk.nKeyCol = 1; 3219 sPk.nColumn = 1; 3220 sPk.aiColumn = &aiColumnPk; 3221 sPk.aiRowLogEst = aiRowEstPk; 3222 sPk.onError = OE_Replace; 3223 sPk.pTable = pTab; 3224 sPk.szIdxRow = pTab->szTabRow; 3225 sPk.idxType = SQLITE_IDXTYPE_IPK; 3226 aiRowEstPk[0] = pTab->nRowLogEst; 3227 aiRowEstPk[1] = 0; 3228 pFirst = pSrc->pTab->pIndex; 3229 if( pSrc->fg.notIndexed==0 ){ 3230 /* The real indices of the table are only considered if the 3231 ** NOT INDEXED qualifier is omitted from the FROM clause */ 3232 sPk.pNext = pFirst; 3233 } 3234 pProbe = &sPk; 3235 } 3236 rSize = pTab->nRowLogEst; 3237 3238 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 3239 /* Automatic indexes */ 3240 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 3241 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 3242 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 3243 && !pSrc->fg.isIndexedBy /* Has no INDEXED BY clause */ 3244 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 3245 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 3246 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 3247 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 3248 ){ 3249 /* Generate auto-index WhereLoops */ 3250 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 3251 WhereTerm *pTerm; 3252 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 3253 rLogSize = estLog(rSize); 3254 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 3255 if( pTerm->prereqRight & pNew->maskSelf ) continue; 3256 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 3257 pNew->u.btree.nEq = 1; 3258 pNew->nSkip = 0; 3259 pNew->u.btree.pIndex = 0; 3260 pNew->nLTerm = 1; 3261 pNew->aLTerm[0] = pTerm; 3262 /* TUNING: One-time cost for computing the automatic index is 3263 ** estimated to be X*N*log2(N) where N is the number of rows in 3264 ** the table being indexed and where X is 7 (LogEst=28) for normal 3265 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value 3266 ** of X is smaller for views and subqueries so that the query planner 3267 ** will be more aggressive about generating automatic indexes for 3268 ** those objects, since there is no opportunity to add schema 3269 ** indexes on subqueries and views. */ 3270 pNew->rSetup = rLogSize + rSize; 3271 if( !IsView(pTab) && (pTab->tabFlags & TF_Ephemeral)==0 ){ 3272 pNew->rSetup += 28; 3273 }else{ 3274 pNew->rSetup -= 10; 3275 } 3276 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 3277 if( pNew->rSetup<0 ) pNew->rSetup = 0; 3278 /* TUNING: Each index lookup yields 20 rows in the table. This 3279 ** is more than the usual guess of 10 rows, since we have no way 3280 ** of knowing how selective the index will ultimately be. It would 3281 ** not be unreasonable to make this value much larger. */ 3282 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 3283 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 3284 pNew->wsFlags = WHERE_AUTO_INDEX; 3285 pNew->prereq = mPrereq | pTerm->prereqRight; 3286 rc = whereLoopInsert(pBuilder, pNew); 3287 } 3288 } 3289 } 3290 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 3291 3292 /* Loop over all indices. If there was an INDEXED BY clause, then only 3293 ** consider index pProbe. */ 3294 for(; rc==SQLITE_OK && pProbe; 3295 pProbe=(pSrc->fg.isIndexedBy ? 0 : pProbe->pNext), iSortIdx++ 3296 ){ 3297 int isLeft = (pSrc->fg.jointype & JT_OUTER)!=0; 3298 if( pProbe->pPartIdxWhere!=0 3299 && !whereUsablePartialIndex(pSrc->iCursor, isLeft, pWC, 3300 pProbe->pPartIdxWhere) 3301 ){ 3302 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 3303 continue; /* Partial index inappropriate for this query */ 3304 } 3305 if( pProbe->bNoQuery ) continue; 3306 rSize = pProbe->aiRowLogEst[0]; 3307 pNew->u.btree.nEq = 0; 3308 pNew->u.btree.nBtm = 0; 3309 pNew->u.btree.nTop = 0; 3310 pNew->nSkip = 0; 3311 pNew->nLTerm = 0; 3312 pNew->iSortIdx = 0; 3313 pNew->rSetup = 0; 3314 pNew->prereq = mPrereq; 3315 pNew->nOut = rSize; 3316 pNew->u.btree.pIndex = pProbe; 3317 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 3318 3319 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 3320 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 3321 if( pProbe->idxType==SQLITE_IDXTYPE_IPK ){ 3322 /* Integer primary key index */ 3323 pNew->wsFlags = WHERE_IPK; 3324 3325 /* Full table scan */ 3326 pNew->iSortIdx = b ? iSortIdx : 0; 3327 /* TUNING: Cost of full table scan is 3.0*N. The 3.0 factor is an 3328 ** extra cost designed to discourage the use of full table scans, 3329 ** since index lookups have better worst-case performance if our 3330 ** stat guesses are wrong. Reduce the 3.0 penalty slightly 3331 ** (to 2.75) if we have valid STAT4 information for the table. 3332 ** At 2.75, a full table scan is preferred over using an index on 3333 ** a column with just two distinct values where each value has about 3334 ** an equal number of appearances. Without STAT4 data, we still want 3335 ** to use an index in that case, since the constraint might be for 3336 ** the scarcer of the two values, and in that case an index lookup is 3337 ** better. 3338 */ 3339 #ifdef SQLITE_ENABLE_STAT4 3340 pNew->rRun = rSize + 16 - 2*((pTab->tabFlags & TF_HasStat4)!=0); 3341 #else 3342 pNew->rRun = rSize + 16; 3343 #endif 3344 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3345 whereLoopOutputAdjust(pWC, pNew, rSize); 3346 rc = whereLoopInsert(pBuilder, pNew); 3347 pNew->nOut = rSize; 3348 if( rc ) break; 3349 }else{ 3350 Bitmask m; 3351 if( pProbe->isCovering ){ 3352 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 3353 m = 0; 3354 }else{ 3355 m = pSrc->colUsed & pProbe->colNotIdxed; 3356 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 3357 } 3358 3359 /* Full scan via index */ 3360 if( b 3361 || !HasRowid(pTab) 3362 || pProbe->pPartIdxWhere!=0 3363 || pSrc->fg.isIndexedBy 3364 || ( m==0 3365 && pProbe->bUnordered==0 3366 && (pProbe->szIdxRow<pTab->szTabRow) 3367 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 3368 && sqlite3GlobalConfig.bUseCis 3369 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 3370 ) 3371 ){ 3372 pNew->iSortIdx = b ? iSortIdx : 0; 3373 3374 /* The cost of visiting the index rows is N*K, where K is 3375 ** between 1.1 and 3.0, depending on the relative sizes of the 3376 ** index and table rows. */ 3377 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 3378 if( m!=0 ){ 3379 /* If this is a non-covering index scan, add in the cost of 3380 ** doing table lookups. The cost will be 3x the number of 3381 ** lookups. Take into account WHERE clause terms that can be 3382 ** satisfied using just the index, and that do not require a 3383 ** table lookup. */ 3384 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 3385 int ii; 3386 int iCur = pSrc->iCursor; 3387 WhereClause *pWC2 = &pWInfo->sWC; 3388 for(ii=0; ii<pWC2->nTerm; ii++){ 3389 WhereTerm *pTerm = &pWC2->a[ii]; 3390 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 3391 break; 3392 } 3393 /* pTerm can be evaluated using just the index. So reduce 3394 ** the expected number of table lookups accordingly */ 3395 if( pTerm->truthProb<=0 ){ 3396 nLookup += pTerm->truthProb; 3397 }else{ 3398 nLookup--; 3399 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 3400 } 3401 } 3402 3403 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 3404 } 3405 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3406 whereLoopOutputAdjust(pWC, pNew, rSize); 3407 rc = whereLoopInsert(pBuilder, pNew); 3408 pNew->nOut = rSize; 3409 if( rc ) break; 3410 } 3411 } 3412 3413 pBuilder->bldFlags1 = 0; 3414 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 3415 if( pBuilder->bldFlags1==SQLITE_BLDF1_INDEXED ){ 3416 /* If a non-unique index is used, or if a prefix of the key for 3417 ** unique index is used (making the index functionally non-unique) 3418 ** then the sqlite_stat1 data becomes important for scoring the 3419 ** plan */ 3420 pTab->tabFlags |= TF_StatsUsed; 3421 } 3422 #ifdef SQLITE_ENABLE_STAT4 3423 sqlite3Stat4ProbeFree(pBuilder->pRec); 3424 pBuilder->nRecValid = 0; 3425 pBuilder->pRec = 0; 3426 #endif 3427 } 3428 return rc; 3429 } 3430 3431 #ifndef SQLITE_OMIT_VIRTUALTABLE 3432 3433 /* 3434 ** Argument pIdxInfo is already populated with all constraints that may 3435 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 3436 ** function marks a subset of those constraints usable, invokes the 3437 ** xBestIndex method and adds the returned plan to pBuilder. 3438 ** 3439 ** A constraint is marked usable if: 3440 ** 3441 ** * Argument mUsable indicates that its prerequisites are available, and 3442 ** 3443 ** * It is not one of the operators specified in the mExclude mask passed 3444 ** as the fourth argument (which in practice is either WO_IN or 0). 3445 ** 3446 ** Argument mPrereq is a mask of tables that must be scanned before the 3447 ** virtual table in question. These are added to the plans prerequisites 3448 ** before it is added to pBuilder. 3449 ** 3450 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3451 ** uses one or more WO_IN terms, or false otherwise. 3452 */ 3453 static int whereLoopAddVirtualOne( 3454 WhereLoopBuilder *pBuilder, 3455 Bitmask mPrereq, /* Mask of tables that must be used. */ 3456 Bitmask mUsable, /* Mask of usable tables */ 3457 u16 mExclude, /* Exclude terms using these operators */ 3458 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3459 u16 mNoOmit, /* Do not omit these constraints */ 3460 int *pbIn /* OUT: True if plan uses an IN(...) op */ 3461 ){ 3462 WhereClause *pWC = pBuilder->pWC; 3463 struct sqlite3_index_constraint *pIdxCons; 3464 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3465 int i; 3466 int mxTerm; 3467 int rc = SQLITE_OK; 3468 WhereLoop *pNew = pBuilder->pNew; 3469 Parse *pParse = pBuilder->pWInfo->pParse; 3470 SrcItem *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3471 int nConstraint = pIdxInfo->nConstraint; 3472 3473 assert( (mUsable & mPrereq)==mPrereq ); 3474 *pbIn = 0; 3475 pNew->prereq = mPrereq; 3476 3477 /* Set the usable flag on the subset of constraints identified by 3478 ** arguments mUsable and mExclude. */ 3479 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3480 for(i=0; i<nConstraint; i++, pIdxCons++){ 3481 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3482 pIdxCons->usable = 0; 3483 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3484 && (pTerm->eOperator & mExclude)==0 3485 ){ 3486 pIdxCons->usable = 1; 3487 } 3488 } 3489 3490 /* Initialize the output fields of the sqlite3_index_info structure */ 3491 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3492 assert( pIdxInfo->needToFreeIdxStr==0 ); 3493 pIdxInfo->idxStr = 0; 3494 pIdxInfo->idxNum = 0; 3495 pIdxInfo->orderByConsumed = 0; 3496 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3497 pIdxInfo->estimatedRows = 25; 3498 pIdxInfo->idxFlags = 0; 3499 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3500 3501 /* Invoke the virtual table xBestIndex() method */ 3502 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3503 if( rc ){ 3504 if( rc==SQLITE_CONSTRAINT ){ 3505 /* If the xBestIndex method returns SQLITE_CONSTRAINT, that means 3506 ** that the particular combination of parameters provided is unusable. 3507 ** Make no entries in the loop table. 3508 */ 3509 WHERETRACE(0xffff, (" ^^^^--- non-viable plan rejected!\n")); 3510 return SQLITE_OK; 3511 } 3512 return rc; 3513 } 3514 3515 mxTerm = -1; 3516 assert( pNew->nLSlot>=nConstraint ); 3517 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 3518 pNew->u.vtab.omitMask = 0; 3519 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3520 for(i=0; i<nConstraint; i++, pIdxCons++){ 3521 int iTerm; 3522 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3523 WhereTerm *pTerm; 3524 int j = pIdxCons->iTermOffset; 3525 if( iTerm>=nConstraint 3526 || j<0 3527 || j>=pWC->nTerm 3528 || pNew->aLTerm[iTerm]!=0 3529 || pIdxCons->usable==0 3530 ){ 3531 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3532 testcase( pIdxInfo->needToFreeIdxStr ); 3533 return SQLITE_ERROR; 3534 } 3535 testcase( iTerm==nConstraint-1 ); 3536 testcase( j==0 ); 3537 testcase( j==pWC->nTerm-1 ); 3538 pTerm = &pWC->a[j]; 3539 pNew->prereq |= pTerm->prereqRight; 3540 assert( iTerm<pNew->nLSlot ); 3541 pNew->aLTerm[iTerm] = pTerm; 3542 if( iTerm>mxTerm ) mxTerm = iTerm; 3543 testcase( iTerm==15 ); 3544 testcase( iTerm==16 ); 3545 if( pUsage[i].omit ){ 3546 if( i<16 && ((1<<i)&mNoOmit)==0 ){ 3547 testcase( i!=iTerm ); 3548 pNew->u.vtab.omitMask |= 1<<iTerm; 3549 }else{ 3550 testcase( i!=iTerm ); 3551 } 3552 } 3553 if( (pTerm->eOperator & WO_IN)!=0 ){ 3554 /* A virtual table that is constrained by an IN clause may not 3555 ** consume the ORDER BY clause because (1) the order of IN terms 3556 ** is not necessarily related to the order of output terms and 3557 ** (2) Multiple outputs from a single IN value will not merge 3558 ** together. */ 3559 pIdxInfo->orderByConsumed = 0; 3560 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3561 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3562 } 3563 } 3564 } 3565 3566 pNew->nLTerm = mxTerm+1; 3567 for(i=0; i<=mxTerm; i++){ 3568 if( pNew->aLTerm[i]==0 ){ 3569 /* The non-zero argvIdx values must be contiguous. Raise an 3570 ** error if they are not */ 3571 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3572 testcase( pIdxInfo->needToFreeIdxStr ); 3573 return SQLITE_ERROR; 3574 } 3575 } 3576 assert( pNew->nLTerm<=pNew->nLSlot ); 3577 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3578 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3579 pIdxInfo->needToFreeIdxStr = 0; 3580 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3581 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3582 pIdxInfo->nOrderBy : 0); 3583 pNew->rSetup = 0; 3584 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3585 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3586 3587 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3588 ** that the scan will visit at most one row. Clear it otherwise. */ 3589 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3590 pNew->wsFlags |= WHERE_ONEROW; 3591 }else{ 3592 pNew->wsFlags &= ~WHERE_ONEROW; 3593 } 3594 rc = whereLoopInsert(pBuilder, pNew); 3595 if( pNew->u.vtab.needFree ){ 3596 sqlite3_free(pNew->u.vtab.idxStr); 3597 pNew->u.vtab.needFree = 0; 3598 } 3599 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3600 *pbIn, (sqlite3_uint64)mPrereq, 3601 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3602 3603 return rc; 3604 } 3605 3606 /* 3607 ** Return the collating sequence for a constraint passed into xBestIndex. 3608 ** 3609 ** pIdxInfo must be an sqlite3_index_info structure passed into xBestIndex. 3610 ** This routine depends on there being a HiddenIndexInfo structure immediately 3611 ** following the sqlite3_index_info structure. 3612 ** 3613 ** Return a pointer to the collation name: 3614 ** 3615 ** 1. If there is an explicit COLLATE operator on the constaint, return it. 3616 ** 3617 ** 2. Else, if the column has an alternative collation, return that. 3618 ** 3619 ** 3. Otherwise, return "BINARY". 3620 */ 3621 const char *sqlite3_vtab_collation(sqlite3_index_info *pIdxInfo, int iCons){ 3622 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3623 const char *zRet = 0; 3624 if( iCons>=0 && iCons<pIdxInfo->nConstraint ){ 3625 CollSeq *pC = 0; 3626 int iTerm = pIdxInfo->aConstraint[iCons].iTermOffset; 3627 Expr *pX = pHidden->pWC->a[iTerm].pExpr; 3628 if( pX->pLeft ){ 3629 pC = sqlite3ExprCompareCollSeq(pHidden->pParse, pX); 3630 } 3631 zRet = (pC ? pC->zName : sqlite3StrBINARY); 3632 } 3633 return zRet; 3634 } 3635 3636 /* 3637 ** Add all WhereLoop objects for a table of the join identified by 3638 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3639 ** 3640 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3641 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3642 ** entries that occur before the virtual table in the FROM clause and are 3643 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3644 ** mUnusable mask contains all FROM clause entries that occur after the 3645 ** virtual table and are separated from it by at least one LEFT or 3646 ** CROSS JOIN. 3647 ** 3648 ** For example, if the query were: 3649 ** 3650 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3651 ** 3652 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3653 ** 3654 ** All the tables in mPrereq must be scanned before the current virtual 3655 ** table. So any terms for which all prerequisites are satisfied by 3656 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3657 ** Conversely, all tables in mUnusable must be scanned after the current 3658 ** virtual table, so any terms for which the prerequisites overlap with 3659 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3660 */ 3661 static int whereLoopAddVirtual( 3662 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3663 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3664 Bitmask mUnusable /* Tables that must be scanned after this one */ 3665 ){ 3666 int rc = SQLITE_OK; /* Return code */ 3667 WhereInfo *pWInfo; /* WHERE analysis context */ 3668 Parse *pParse; /* The parsing context */ 3669 WhereClause *pWC; /* The WHERE clause */ 3670 SrcItem *pSrc; /* The FROM clause term to search */ 3671 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3672 int nConstraint; /* Number of constraints in p */ 3673 int bIn; /* True if plan uses IN(...) operator */ 3674 WhereLoop *pNew; 3675 Bitmask mBest; /* Tables used by best possible plan */ 3676 u16 mNoOmit; 3677 3678 assert( (mPrereq & mUnusable)==0 ); 3679 pWInfo = pBuilder->pWInfo; 3680 pParse = pWInfo->pParse; 3681 pWC = pBuilder->pWC; 3682 pNew = pBuilder->pNew; 3683 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3684 assert( IsVirtual(pSrc->pTab) ); 3685 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy, 3686 &mNoOmit); 3687 if( p==0 ) return SQLITE_NOMEM_BKPT; 3688 pNew->rSetup = 0; 3689 pNew->wsFlags = WHERE_VIRTUALTABLE; 3690 pNew->nLTerm = 0; 3691 pNew->u.vtab.needFree = 0; 3692 nConstraint = p->nConstraint; 3693 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3694 sqlite3DbFree(pParse->db, p); 3695 return SQLITE_NOMEM_BKPT; 3696 } 3697 3698 /* First call xBestIndex() with all constraints usable. */ 3699 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc->pTab->zName)); 3700 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3701 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn); 3702 3703 /* If the call to xBestIndex() with all terms enabled produced a plan 3704 ** that does not require any source tables (IOW: a plan with mBest==0) 3705 ** and does not use an IN(...) operator, then there is no point in making 3706 ** any further calls to xBestIndex() since they will all return the same 3707 ** result (if the xBestIndex() implementation is sane). */ 3708 if( rc==SQLITE_OK && ((mBest = (pNew->prereq & ~mPrereq))!=0 || bIn) ){ 3709 int seenZero = 0; /* True if a plan with no prereqs seen */ 3710 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3711 Bitmask mPrev = 0; 3712 Bitmask mBestNoIn = 0; 3713 3714 /* If the plan produced by the earlier call uses an IN(...) term, call 3715 ** xBestIndex again, this time with IN(...) terms disabled. */ 3716 if( bIn ){ 3717 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3718 rc = whereLoopAddVirtualOne( 3719 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn); 3720 assert( bIn==0 ); 3721 mBestNoIn = pNew->prereq & ~mPrereq; 3722 if( mBestNoIn==0 ){ 3723 seenZero = 1; 3724 seenZeroNoIN = 1; 3725 } 3726 } 3727 3728 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3729 ** in the set of terms that apply to the current virtual table. */ 3730 while( rc==SQLITE_OK ){ 3731 int i; 3732 Bitmask mNext = ALLBITS; 3733 assert( mNext>0 ); 3734 for(i=0; i<nConstraint; i++){ 3735 Bitmask mThis = ( 3736 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3737 ); 3738 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3739 } 3740 mPrev = mNext; 3741 if( mNext==ALLBITS ) break; 3742 if( mNext==mBest || mNext==mBestNoIn ) continue; 3743 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3744 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3745 rc = whereLoopAddVirtualOne( 3746 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn); 3747 if( pNew->prereq==mPrereq ){ 3748 seenZero = 1; 3749 if( bIn==0 ) seenZeroNoIN = 1; 3750 } 3751 } 3752 3753 /* If the calls to xBestIndex() in the above loop did not find a plan 3754 ** that requires no source tables at all (i.e. one guaranteed to be 3755 ** usable), make a call here with all source tables disabled */ 3756 if( rc==SQLITE_OK && seenZero==0 ){ 3757 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3758 rc = whereLoopAddVirtualOne( 3759 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn); 3760 if( bIn==0 ) seenZeroNoIN = 1; 3761 } 3762 3763 /* If the calls to xBestIndex() have so far failed to find a plan 3764 ** that requires no source tables at all and does not use an IN(...) 3765 ** operator, make a final call to obtain one here. */ 3766 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3767 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3768 rc = whereLoopAddVirtualOne( 3769 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn); 3770 } 3771 } 3772 3773 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3774 sqlite3DbFreeNN(pParse->db, p); 3775 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc->pTab->zName, rc)); 3776 return rc; 3777 } 3778 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3779 3780 /* 3781 ** Add WhereLoop entries to handle OR terms. This works for either 3782 ** btrees or virtual tables. 3783 */ 3784 static int whereLoopAddOr( 3785 WhereLoopBuilder *pBuilder, 3786 Bitmask mPrereq, 3787 Bitmask mUnusable 3788 ){ 3789 WhereInfo *pWInfo = pBuilder->pWInfo; 3790 WhereClause *pWC; 3791 WhereLoop *pNew; 3792 WhereTerm *pTerm, *pWCEnd; 3793 int rc = SQLITE_OK; 3794 int iCur; 3795 WhereClause tempWC; 3796 WhereLoopBuilder sSubBuild; 3797 WhereOrSet sSum, sCur; 3798 SrcItem *pItem; 3799 3800 pWC = pBuilder->pWC; 3801 pWCEnd = pWC->a + pWC->nTerm; 3802 pNew = pBuilder->pNew; 3803 memset(&sSum, 0, sizeof(sSum)); 3804 pItem = pWInfo->pTabList->a + pNew->iTab; 3805 iCur = pItem->iCursor; 3806 3807 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3808 if( (pTerm->eOperator & WO_OR)!=0 3809 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3810 ){ 3811 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3812 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3813 WhereTerm *pOrTerm; 3814 int once = 1; 3815 int i, j; 3816 3817 sSubBuild = *pBuilder; 3818 sSubBuild.pOrderBy = 0; 3819 sSubBuild.pOrSet = &sCur; 3820 3821 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3822 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3823 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3824 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3825 }else if( pOrTerm->leftCursor==iCur ){ 3826 tempWC.pWInfo = pWC->pWInfo; 3827 tempWC.pOuter = pWC; 3828 tempWC.op = TK_AND; 3829 tempWC.nTerm = 1; 3830 tempWC.nBase = 1; 3831 tempWC.a = pOrTerm; 3832 sSubBuild.pWC = &tempWC; 3833 }else{ 3834 continue; 3835 } 3836 sCur.n = 0; 3837 #ifdef WHERETRACE_ENABLED 3838 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3839 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3840 if( sqlite3WhereTrace & 0x400 ){ 3841 sqlite3WhereClausePrint(sSubBuild.pWC); 3842 } 3843 #endif 3844 #ifndef SQLITE_OMIT_VIRTUALTABLE 3845 if( IsVirtual(pItem->pTab) ){ 3846 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 3847 }else 3848 #endif 3849 { 3850 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 3851 } 3852 if( rc==SQLITE_OK ){ 3853 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 3854 } 3855 assert( rc==SQLITE_OK || rc==SQLITE_DONE || sCur.n==0 3856 || rc==SQLITE_NOMEM ); 3857 testcase( rc==SQLITE_NOMEM && sCur.n>0 ); 3858 testcase( rc==SQLITE_DONE ); 3859 if( sCur.n==0 ){ 3860 sSum.n = 0; 3861 break; 3862 }else if( once ){ 3863 whereOrMove(&sSum, &sCur); 3864 once = 0; 3865 }else{ 3866 WhereOrSet sPrev; 3867 whereOrMove(&sPrev, &sSum); 3868 sSum.n = 0; 3869 for(i=0; i<sPrev.n; i++){ 3870 for(j=0; j<sCur.n; j++){ 3871 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3872 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3873 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3874 } 3875 } 3876 } 3877 } 3878 pNew->nLTerm = 1; 3879 pNew->aLTerm[0] = pTerm; 3880 pNew->wsFlags = WHERE_MULTI_OR; 3881 pNew->rSetup = 0; 3882 pNew->iSortIdx = 0; 3883 memset(&pNew->u, 0, sizeof(pNew->u)); 3884 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3885 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3886 ** of all sub-scans required by the OR-scan. However, due to rounding 3887 ** errors, it may be that the cost of the OR-scan is equal to its 3888 ** most expensive sub-scan. Add the smallest possible penalty 3889 ** (equivalent to multiplying the cost by 1.07) to ensure that 3890 ** this does not happen. Otherwise, for WHERE clauses such as the 3891 ** following where there is an index on "y": 3892 ** 3893 ** WHERE likelihood(x=?, 0.99) OR y=? 3894 ** 3895 ** the planner may elect to "OR" together a full-table scan and an 3896 ** index lookup. And other similarly odd results. */ 3897 pNew->rRun = sSum.a[i].rRun + 1; 3898 pNew->nOut = sSum.a[i].nOut; 3899 pNew->prereq = sSum.a[i].prereq; 3900 rc = whereLoopInsert(pBuilder, pNew); 3901 } 3902 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3903 } 3904 } 3905 return rc; 3906 } 3907 3908 /* 3909 ** Add all WhereLoop objects for all tables 3910 */ 3911 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3912 WhereInfo *pWInfo = pBuilder->pWInfo; 3913 Bitmask mPrereq = 0; 3914 Bitmask mPrior = 0; 3915 int iTab; 3916 SrcList *pTabList = pWInfo->pTabList; 3917 SrcItem *pItem; 3918 SrcItem *pEnd = &pTabList->a[pWInfo->nLevel]; 3919 sqlite3 *db = pWInfo->pParse->db; 3920 int rc = SQLITE_OK; 3921 WhereLoop *pNew; 3922 3923 /* Loop over the tables in the join, from left to right */ 3924 pNew = pBuilder->pNew; 3925 whereLoopInit(pNew); 3926 pBuilder->iPlanLimit = SQLITE_QUERY_PLANNER_LIMIT; 3927 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3928 Bitmask mUnusable = 0; 3929 pNew->iTab = iTab; 3930 pBuilder->iPlanLimit += SQLITE_QUERY_PLANNER_LIMIT_INCR; 3931 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3932 if( (pItem->fg.jointype & (JT_LEFT|JT_CROSS))!=0 ){ 3933 /* This condition is true when pItem is the FROM clause term on the 3934 ** right-hand-side of a LEFT or CROSS JOIN. */ 3935 mPrereq = mPrior; 3936 }else{ 3937 mPrereq = 0; 3938 } 3939 #ifndef SQLITE_OMIT_VIRTUALTABLE 3940 if( IsVirtual(pItem->pTab) ){ 3941 SrcItem *p; 3942 for(p=&pItem[1]; p<pEnd; p++){ 3943 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3944 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3945 } 3946 } 3947 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 3948 }else 3949 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3950 { 3951 rc = whereLoopAddBtree(pBuilder, mPrereq); 3952 } 3953 if( rc==SQLITE_OK && pBuilder->pWC->hasOr ){ 3954 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 3955 } 3956 mPrior |= pNew->maskSelf; 3957 if( rc || db->mallocFailed ){ 3958 if( rc==SQLITE_DONE ){ 3959 /* We hit the query planner search limit set by iPlanLimit */ 3960 sqlite3_log(SQLITE_WARNING, "abbreviated query algorithm search"); 3961 rc = SQLITE_OK; 3962 }else{ 3963 break; 3964 } 3965 } 3966 } 3967 3968 whereLoopClear(db, pNew); 3969 return rc; 3970 } 3971 3972 /* 3973 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 3974 ** parameters) to see if it outputs rows in the requested ORDER BY 3975 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3976 ** 3977 ** N>0: N terms of the ORDER BY clause are satisfied 3978 ** N==0: No terms of the ORDER BY clause are satisfied 3979 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3980 ** 3981 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3982 ** strict. With GROUP BY and DISTINCT the only requirement is that 3983 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3984 ** and DISTINCT do not require rows to appear in any particular order as long 3985 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3986 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3987 ** pOrderBy terms must be matched in strict left-to-right order. 3988 */ 3989 static i8 wherePathSatisfiesOrderBy( 3990 WhereInfo *pWInfo, /* The WHERE clause */ 3991 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3992 WherePath *pPath, /* The WherePath to check */ 3993 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 3994 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3995 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3996 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3997 ){ 3998 u8 revSet; /* True if rev is known */ 3999 u8 rev; /* Composite sort order */ 4000 u8 revIdx; /* Index sort order */ 4001 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 4002 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 4003 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 4004 u16 eqOpMask; /* Allowed equality operators */ 4005 u16 nKeyCol; /* Number of key columns in pIndex */ 4006 u16 nColumn; /* Total number of ordered columns in the index */ 4007 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 4008 int iLoop; /* Index of WhereLoop in pPath being processed */ 4009 int i, j; /* Loop counters */ 4010 int iCur; /* Cursor number for current WhereLoop */ 4011 int iColumn; /* A column number within table iCur */ 4012 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 4013 WhereTerm *pTerm; /* A single term of the WHERE clause */ 4014 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 4015 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 4016 Index *pIndex; /* The index associated with pLoop */ 4017 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 4018 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 4019 Bitmask obDone; /* Mask of all ORDER BY terms */ 4020 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 4021 Bitmask ready; /* Mask of inner loops */ 4022 4023 /* 4024 ** We say the WhereLoop is "one-row" if it generates no more than one 4025 ** row of output. A WhereLoop is one-row if all of the following are true: 4026 ** (a) All index columns match with WHERE_COLUMN_EQ. 4027 ** (b) The index is unique 4028 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 4029 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 4030 ** 4031 ** We say the WhereLoop is "order-distinct" if the set of columns from 4032 ** that WhereLoop that are in the ORDER BY clause are different for every 4033 ** row of the WhereLoop. Every one-row WhereLoop is automatically 4034 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 4035 ** is not order-distinct. To be order-distinct is not quite the same as being 4036 ** UNIQUE since a UNIQUE column or index can have multiple rows that 4037 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 4038 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 4039 ** 4040 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 4041 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 4042 ** automatically order-distinct. 4043 */ 4044 4045 assert( pOrderBy!=0 ); 4046 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 4047 4048 nOrderBy = pOrderBy->nExpr; 4049 testcase( nOrderBy==BMS-1 ); 4050 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 4051 isOrderDistinct = 1; 4052 obDone = MASKBIT(nOrderBy)-1; 4053 orderDistinctMask = 0; 4054 ready = 0; 4055 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 4056 if( wctrlFlags & (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MAX|WHERE_ORDERBY_MIN) ){ 4057 eqOpMask |= WO_IN; 4058 } 4059 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 4060 if( iLoop>0 ) ready |= pLoop->maskSelf; 4061 if( iLoop<nLoop ){ 4062 pLoop = pPath->aLoop[iLoop]; 4063 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 4064 }else{ 4065 pLoop = pLast; 4066 } 4067 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 4068 if( pLoop->u.vtab.isOrdered && (wctrlFlags & WHERE_DISTINCTBY)==0 ){ 4069 obSat = obDone; 4070 } 4071 break; 4072 }else if( wctrlFlags & WHERE_DISTINCTBY ){ 4073 pLoop->u.btree.nDistinctCol = 0; 4074 } 4075 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 4076 4077 /* Mark off any ORDER BY term X that is a column in the table of 4078 ** the current loop for which there is term in the WHERE 4079 ** clause of the form X IS NULL or X=? that reference only outer 4080 ** loops. 4081 */ 4082 for(i=0; i<nOrderBy; i++){ 4083 if( MASKBIT(i) & obSat ) continue; 4084 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr); 4085 if( NEVER(pOBExpr==0) ) continue; 4086 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue; 4087 if( pOBExpr->iTable!=iCur ) continue; 4088 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 4089 ~ready, eqOpMask, 0); 4090 if( pTerm==0 ) continue; 4091 if( pTerm->eOperator==WO_IN ){ 4092 /* IN terms are only valid for sorting in the ORDER BY LIMIT 4093 ** optimization, and then only if they are actually used 4094 ** by the query plan */ 4095 assert( wctrlFlags & 4096 (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX) ); 4097 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 4098 if( j>=pLoop->nLTerm ) continue; 4099 } 4100 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 4101 Parse *pParse = pWInfo->pParse; 4102 CollSeq *pColl1 = sqlite3ExprNNCollSeq(pParse, pOrderBy->a[i].pExpr); 4103 CollSeq *pColl2 = sqlite3ExprCompareCollSeq(pParse, pTerm->pExpr); 4104 assert( pColl1 ); 4105 if( pColl2==0 || sqlite3StrICmp(pColl1->zName, pColl2->zName) ){ 4106 continue; 4107 } 4108 testcase( pTerm->pExpr->op==TK_IS ); 4109 } 4110 obSat |= MASKBIT(i); 4111 } 4112 4113 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 4114 if( pLoop->wsFlags & WHERE_IPK ){ 4115 pIndex = 0; 4116 nKeyCol = 0; 4117 nColumn = 1; 4118 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 4119 return 0; 4120 }else{ 4121 nKeyCol = pIndex->nKeyCol; 4122 nColumn = pIndex->nColumn; 4123 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 4124 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 4125 || !HasRowid(pIndex->pTable)); 4126 /* All relevant terms of the index must also be non-NULL in order 4127 ** for isOrderDistinct to be true. So the isOrderDistint value 4128 ** computed here might be a false positive. Corrections will be 4129 ** made at tag-20210426-1 below */ 4130 isOrderDistinct = IsUniqueIndex(pIndex) 4131 && (pLoop->wsFlags & WHERE_SKIPSCAN)==0; 4132 } 4133 4134 /* Loop through all columns of the index and deal with the ones 4135 ** that are not constrained by == or IN. 4136 */ 4137 rev = revSet = 0; 4138 distinctColumns = 0; 4139 for(j=0; j<nColumn; j++){ 4140 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 4141 4142 assert( j>=pLoop->u.btree.nEq 4143 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 4144 ); 4145 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 4146 u16 eOp = pLoop->aLTerm[j]->eOperator; 4147 4148 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 4149 ** doing WHERE_ORDERBY_LIMIT processing). Except, IS and ISNULL 4150 ** terms imply that the index is not UNIQUE NOT NULL in which case 4151 ** the loop need to be marked as not order-distinct because it can 4152 ** have repeated NULL rows. 4153 ** 4154 ** If the current term is a column of an ((?,?) IN (SELECT...)) 4155 ** expression for which the SELECT returns more than one column, 4156 ** check that it is the only column used by this loop. Otherwise, 4157 ** if it is one of two or more, none of the columns can be 4158 ** considered to match an ORDER BY term. 4159 */ 4160 if( (eOp & eqOpMask)!=0 ){ 4161 if( eOp & (WO_ISNULL|WO_IS) ){ 4162 testcase( eOp & WO_ISNULL ); 4163 testcase( eOp & WO_IS ); 4164 testcase( isOrderDistinct ); 4165 isOrderDistinct = 0; 4166 } 4167 continue; 4168 }else if( ALWAYS(eOp & WO_IN) ){ 4169 /* ALWAYS() justification: eOp is an equality operator due to the 4170 ** j<pLoop->u.btree.nEq constraint above. Any equality other 4171 ** than WO_IN is captured by the previous "if". So this one 4172 ** always has to be WO_IN. */ 4173 Expr *pX = pLoop->aLTerm[j]->pExpr; 4174 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 4175 if( pLoop->aLTerm[i]->pExpr==pX ){ 4176 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 4177 bOnce = 0; 4178 break; 4179 } 4180 } 4181 } 4182 } 4183 4184 /* Get the column number in the table (iColumn) and sort order 4185 ** (revIdx) for the j-th column of the index. 4186 */ 4187 if( pIndex ){ 4188 iColumn = pIndex->aiColumn[j]; 4189 revIdx = pIndex->aSortOrder[j] & KEYINFO_ORDER_DESC; 4190 if( iColumn==pIndex->pTable->iPKey ) iColumn = XN_ROWID; 4191 }else{ 4192 iColumn = XN_ROWID; 4193 revIdx = 0; 4194 } 4195 4196 /* An unconstrained column that might be NULL means that this 4197 ** WhereLoop is not well-ordered. tag-20210426-1 4198 */ 4199 if( isOrderDistinct ){ 4200 if( iColumn>=0 4201 && j>=pLoop->u.btree.nEq 4202 && pIndex->pTable->aCol[iColumn].notNull==0 4203 ){ 4204 isOrderDistinct = 0; 4205 } 4206 if( iColumn==XN_EXPR ){ 4207 isOrderDistinct = 0; 4208 } 4209 } 4210 4211 /* Find the ORDER BY term that corresponds to the j-th column 4212 ** of the index and mark that ORDER BY term off 4213 */ 4214 isMatch = 0; 4215 for(i=0; bOnce && i<nOrderBy; i++){ 4216 if( MASKBIT(i) & obSat ) continue; 4217 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr); 4218 testcase( wctrlFlags & WHERE_GROUPBY ); 4219 testcase( wctrlFlags & WHERE_DISTINCTBY ); 4220 if( NEVER(pOBExpr==0) ) continue; 4221 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 4222 if( iColumn>=XN_ROWID ){ 4223 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue; 4224 if( pOBExpr->iTable!=iCur ) continue; 4225 if( pOBExpr->iColumn!=iColumn ) continue; 4226 }else{ 4227 Expr *pIdxExpr = pIndex->aColExpr->a[j].pExpr; 4228 if( sqlite3ExprCompareSkip(pOBExpr, pIdxExpr, iCur) ){ 4229 continue; 4230 } 4231 } 4232 if( iColumn!=XN_ROWID ){ 4233 pColl = sqlite3ExprNNCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 4234 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 4235 } 4236 if( wctrlFlags & WHERE_DISTINCTBY ){ 4237 pLoop->u.btree.nDistinctCol = j+1; 4238 } 4239 isMatch = 1; 4240 break; 4241 } 4242 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 4243 /* Make sure the sort order is compatible in an ORDER BY clause. 4244 ** Sort order is irrelevant for a GROUP BY clause. */ 4245 if( revSet ){ 4246 if( (rev ^ revIdx)!=(pOrderBy->a[i].sortFlags&KEYINFO_ORDER_DESC) ){ 4247 isMatch = 0; 4248 } 4249 }else{ 4250 rev = revIdx ^ (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC); 4251 if( rev ) *pRevMask |= MASKBIT(iLoop); 4252 revSet = 1; 4253 } 4254 } 4255 if( isMatch && (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL) ){ 4256 if( j==pLoop->u.btree.nEq ){ 4257 pLoop->wsFlags |= WHERE_BIGNULL_SORT; 4258 }else{ 4259 isMatch = 0; 4260 } 4261 } 4262 if( isMatch ){ 4263 if( iColumn==XN_ROWID ){ 4264 testcase( distinctColumns==0 ); 4265 distinctColumns = 1; 4266 } 4267 obSat |= MASKBIT(i); 4268 }else{ 4269 /* No match found */ 4270 if( j==0 || j<nKeyCol ){ 4271 testcase( isOrderDistinct!=0 ); 4272 isOrderDistinct = 0; 4273 } 4274 break; 4275 } 4276 } /* end Loop over all index columns */ 4277 if( distinctColumns ){ 4278 testcase( isOrderDistinct==0 ); 4279 isOrderDistinct = 1; 4280 } 4281 } /* end-if not one-row */ 4282 4283 /* Mark off any other ORDER BY terms that reference pLoop */ 4284 if( isOrderDistinct ){ 4285 orderDistinctMask |= pLoop->maskSelf; 4286 for(i=0; i<nOrderBy; i++){ 4287 Expr *p; 4288 Bitmask mTerm; 4289 if( MASKBIT(i) & obSat ) continue; 4290 p = pOrderBy->a[i].pExpr; 4291 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 4292 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 4293 if( (mTerm&~orderDistinctMask)==0 ){ 4294 obSat |= MASKBIT(i); 4295 } 4296 } 4297 } 4298 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 4299 if( obSat==obDone ) return (i8)nOrderBy; 4300 if( !isOrderDistinct ){ 4301 for(i=nOrderBy-1; i>0; i--){ 4302 Bitmask m = ALWAYS(i<BMS) ? MASKBIT(i) - 1 : 0; 4303 if( (obSat&m)==m ) return i; 4304 } 4305 return 0; 4306 } 4307 return -1; 4308 } 4309 4310 4311 /* 4312 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 4313 ** the planner assumes that the specified pOrderBy list is actually a GROUP 4314 ** BY clause - and so any order that groups rows as required satisfies the 4315 ** request. 4316 ** 4317 ** Normally, in this case it is not possible for the caller to determine 4318 ** whether or not the rows are really being delivered in sorted order, or 4319 ** just in some other order that provides the required grouping. However, 4320 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 4321 ** this function may be called on the returned WhereInfo object. It returns 4322 ** true if the rows really will be sorted in the specified order, or false 4323 ** otherwise. 4324 ** 4325 ** For example, assuming: 4326 ** 4327 ** CREATE INDEX i1 ON t1(x, Y); 4328 ** 4329 ** then 4330 ** 4331 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 4332 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 4333 */ 4334 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 4335 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 4336 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 4337 return pWInfo->sorted; 4338 } 4339 4340 #ifdef WHERETRACE_ENABLED 4341 /* For debugging use only: */ 4342 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 4343 static char zName[65]; 4344 int i; 4345 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 4346 if( pLast ) zName[i++] = pLast->cId; 4347 zName[i] = 0; 4348 return zName; 4349 } 4350 #endif 4351 4352 /* 4353 ** Return the cost of sorting nRow rows, assuming that the keys have 4354 ** nOrderby columns and that the first nSorted columns are already in 4355 ** order. 4356 */ 4357 static LogEst whereSortingCost( 4358 WhereInfo *pWInfo, 4359 LogEst nRow, 4360 int nOrderBy, 4361 int nSorted 4362 ){ 4363 /* TUNING: Estimated cost of a full external sort, where N is 4364 ** the number of rows to sort is: 4365 ** 4366 ** cost = (3.0 * N * log(N)). 4367 ** 4368 ** Or, if the order-by clause has X terms but only the last Y 4369 ** terms are out of order, then block-sorting will reduce the 4370 ** sorting cost to: 4371 ** 4372 ** cost = (3.0 * N * log(N)) * (Y/X) 4373 ** 4374 ** The (Y/X) term is implemented using stack variable rScale 4375 ** below. 4376 */ 4377 LogEst rScale, rSortCost; 4378 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 4379 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 4380 rSortCost = nRow + rScale + 16; 4381 4382 /* Multiple by log(M) where M is the number of output rows. 4383 ** Use the LIMIT for M if it is smaller. Or if this sort is for 4384 ** a DISTINCT operator, M will be the number of distinct output 4385 ** rows, so fudge it downwards a bit. 4386 */ 4387 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 4388 nRow = pWInfo->iLimit; 4389 }else if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT) ){ 4390 /* TUNING: In the sort for a DISTINCT operator, assume that the DISTINCT 4391 ** reduces the number of output rows by a factor of 2 */ 4392 if( nRow>10 ){ nRow -= 10; assert( 10==sqlite3LogEst(2) ); } 4393 } 4394 rSortCost += estLog(nRow); 4395 return rSortCost; 4396 } 4397 4398 /* 4399 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 4400 ** attempts to find the lowest cost path that visits each WhereLoop 4401 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 4402 ** 4403 ** Assume that the total number of output rows that will need to be sorted 4404 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 4405 ** costs if nRowEst==0. 4406 ** 4407 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 4408 ** error occurs. 4409 */ 4410 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 4411 int mxChoice; /* Maximum number of simultaneous paths tracked */ 4412 int nLoop; /* Number of terms in the join */ 4413 Parse *pParse; /* Parsing context */ 4414 sqlite3 *db; /* The database connection */ 4415 int iLoop; /* Loop counter over the terms of the join */ 4416 int ii, jj; /* Loop counters */ 4417 int mxI = 0; /* Index of next entry to replace */ 4418 int nOrderBy; /* Number of ORDER BY clause terms */ 4419 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 4420 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 4421 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 4422 WherePath *aFrom; /* All nFrom paths at the previous level */ 4423 WherePath *aTo; /* The nTo best paths at the current level */ 4424 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 4425 WherePath *pTo; /* An element of aTo[] that we are working on */ 4426 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 4427 WhereLoop **pX; /* Used to divy up the pSpace memory */ 4428 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 4429 char *pSpace; /* Temporary memory used by this routine */ 4430 int nSpace; /* Bytes of space allocated at pSpace */ 4431 4432 pParse = pWInfo->pParse; 4433 db = pParse->db; 4434 nLoop = pWInfo->nLevel; 4435 /* TUNING: For simple queries, only the best path is tracked. 4436 ** For 2-way joins, the 5 best paths are followed. 4437 ** For joins of 3 or more tables, track the 10 best paths */ 4438 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 4439 assert( nLoop<=pWInfo->pTabList->nSrc ); 4440 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 4441 4442 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 4443 ** case the purpose of this call is to estimate the number of rows returned 4444 ** by the overall query. Once this estimate has been obtained, the caller 4445 ** will invoke this function a second time, passing the estimate as the 4446 ** nRowEst parameter. */ 4447 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 4448 nOrderBy = 0; 4449 }else{ 4450 nOrderBy = pWInfo->pOrderBy->nExpr; 4451 } 4452 4453 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 4454 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 4455 nSpace += sizeof(LogEst) * nOrderBy; 4456 pSpace = sqlite3DbMallocRawNN(db, nSpace); 4457 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 4458 aTo = (WherePath*)pSpace; 4459 aFrom = aTo+mxChoice; 4460 memset(aFrom, 0, sizeof(aFrom[0])); 4461 pX = (WhereLoop**)(aFrom+mxChoice); 4462 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 4463 pFrom->aLoop = pX; 4464 } 4465 if( nOrderBy ){ 4466 /* If there is an ORDER BY clause and it is not being ignored, set up 4467 ** space for the aSortCost[] array. Each element of the aSortCost array 4468 ** is either zero - meaning it has not yet been initialized - or the 4469 ** cost of sorting nRowEst rows of data where the first X terms of 4470 ** the ORDER BY clause are already in order, where X is the array 4471 ** index. */ 4472 aSortCost = (LogEst*)pX; 4473 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 4474 } 4475 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 4476 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 4477 4478 /* Seed the search with a single WherePath containing zero WhereLoops. 4479 ** 4480 ** TUNING: Do not let the number of iterations go above 28. If the cost 4481 ** of computing an automatic index is not paid back within the first 28 4482 ** rows, then do not use the automatic index. */ 4483 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 4484 nFrom = 1; 4485 assert( aFrom[0].isOrdered==0 ); 4486 if( nOrderBy ){ 4487 /* If nLoop is zero, then there are no FROM terms in the query. Since 4488 ** in this case the query may return a maximum of one row, the results 4489 ** are already in the requested order. Set isOrdered to nOrderBy to 4490 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 4491 ** -1, indicating that the result set may or may not be ordered, 4492 ** depending on the loops added to the current plan. */ 4493 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 4494 } 4495 4496 /* Compute successively longer WherePaths using the previous generation 4497 ** of WherePaths as the basis for the next. Keep track of the mxChoice 4498 ** best paths at each generation */ 4499 for(iLoop=0; iLoop<nLoop; iLoop++){ 4500 nTo = 0; 4501 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 4502 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 4503 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 4504 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 4505 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 4506 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 4507 Bitmask maskNew; /* Mask of src visited by (..) */ 4508 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 4509 4510 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 4511 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 4512 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<3 ){ 4513 /* Do not use an automatic index if the this loop is expected 4514 ** to run less than 1.25 times. It is tempting to also exclude 4515 ** automatic index usage on an outer loop, but sometimes an automatic 4516 ** index is useful in the outer loop of a correlated subquery. */ 4517 assert( 10==sqlite3LogEst(2) ); 4518 continue; 4519 } 4520 4521 /* At this point, pWLoop is a candidate to be the next loop. 4522 ** Compute its cost */ 4523 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 4524 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 4525 nOut = pFrom->nRow + pWLoop->nOut; 4526 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 4527 if( isOrdered<0 ){ 4528 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 4529 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 4530 iLoop, pWLoop, &revMask); 4531 }else{ 4532 revMask = pFrom->revLoop; 4533 } 4534 if( isOrdered>=0 && isOrdered<nOrderBy ){ 4535 if( aSortCost[isOrdered]==0 ){ 4536 aSortCost[isOrdered] = whereSortingCost( 4537 pWInfo, nRowEst, nOrderBy, isOrdered 4538 ); 4539 } 4540 /* TUNING: Add a small extra penalty (5) to sorting as an 4541 ** extra encouragment to the query planner to select a plan 4542 ** where the rows emerge in the correct order without any sorting 4543 ** required. */ 4544 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]) + 5; 4545 4546 WHERETRACE(0x002, 4547 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 4548 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 4549 rUnsorted, rCost)); 4550 }else{ 4551 rCost = rUnsorted; 4552 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 4553 } 4554 4555 /* Check to see if pWLoop should be added to the set of 4556 ** mxChoice best-so-far paths. 4557 ** 4558 ** First look for an existing path among best-so-far paths 4559 ** that covers the same set of loops and has the same isOrdered 4560 ** setting as the current path candidate. 4561 ** 4562 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 4563 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 4564 ** of legal values for isOrdered, -1..64. 4565 */ 4566 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4567 if( pTo->maskLoop==maskNew 4568 && ((pTo->isOrdered^isOrdered)&0x80)==0 4569 ){ 4570 testcase( jj==nTo-1 ); 4571 break; 4572 } 4573 } 4574 if( jj>=nTo ){ 4575 /* None of the existing best-so-far paths match the candidate. */ 4576 if( nTo>=mxChoice 4577 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4578 ){ 4579 /* The current candidate is no better than any of the mxChoice 4580 ** paths currently in the best-so-far buffer. So discard 4581 ** this candidate as not viable. */ 4582 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4583 if( sqlite3WhereTrace&0x4 ){ 4584 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4585 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4586 isOrdered>=0 ? isOrdered+'0' : '?'); 4587 } 4588 #endif 4589 continue; 4590 } 4591 /* If we reach this points it means that the new candidate path 4592 ** needs to be added to the set of best-so-far paths. */ 4593 if( nTo<mxChoice ){ 4594 /* Increase the size of the aTo set by one */ 4595 jj = nTo++; 4596 }else{ 4597 /* New path replaces the prior worst to keep count below mxChoice */ 4598 jj = mxI; 4599 } 4600 pTo = &aTo[jj]; 4601 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4602 if( sqlite3WhereTrace&0x4 ){ 4603 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4604 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4605 isOrdered>=0 ? isOrdered+'0' : '?'); 4606 } 4607 #endif 4608 }else{ 4609 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4610 ** same set of loops and has the same isOrdered setting as the 4611 ** candidate path. Check to see if the candidate should replace 4612 ** pTo or if the candidate should be skipped. 4613 ** 4614 ** The conditional is an expanded vector comparison equivalent to: 4615 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4616 */ 4617 if( pTo->rCost<rCost 4618 || (pTo->rCost==rCost 4619 && (pTo->nRow<nOut 4620 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4621 ) 4622 ) 4623 ){ 4624 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4625 if( sqlite3WhereTrace&0x4 ){ 4626 sqlite3DebugPrintf( 4627 "Skip %s cost=%-3d,%3d,%3d order=%c", 4628 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4629 isOrdered>=0 ? isOrdered+'0' : '?'); 4630 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4631 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4632 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4633 } 4634 #endif 4635 /* Discard the candidate path from further consideration */ 4636 testcase( pTo->rCost==rCost ); 4637 continue; 4638 } 4639 testcase( pTo->rCost==rCost+1 ); 4640 /* Control reaches here if the candidate path is better than the 4641 ** pTo path. Replace pTo with the candidate. */ 4642 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4643 if( sqlite3WhereTrace&0x4 ){ 4644 sqlite3DebugPrintf( 4645 "Update %s cost=%-3d,%3d,%3d order=%c", 4646 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4647 isOrdered>=0 ? isOrdered+'0' : '?'); 4648 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4649 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4650 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4651 } 4652 #endif 4653 } 4654 /* pWLoop is a winner. Add it to the set of best so far */ 4655 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4656 pTo->revLoop = revMask; 4657 pTo->nRow = nOut; 4658 pTo->rCost = rCost; 4659 pTo->rUnsorted = rUnsorted; 4660 pTo->isOrdered = isOrdered; 4661 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4662 pTo->aLoop[iLoop] = pWLoop; 4663 if( nTo>=mxChoice ){ 4664 mxI = 0; 4665 mxCost = aTo[0].rCost; 4666 mxUnsorted = aTo[0].nRow; 4667 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4668 if( pTo->rCost>mxCost 4669 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4670 ){ 4671 mxCost = pTo->rCost; 4672 mxUnsorted = pTo->rUnsorted; 4673 mxI = jj; 4674 } 4675 } 4676 } 4677 } 4678 } 4679 4680 #ifdef WHERETRACE_ENABLED /* >=2 */ 4681 if( sqlite3WhereTrace & 0x02 ){ 4682 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4683 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4684 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4685 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4686 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4687 if( pTo->isOrdered>0 ){ 4688 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4689 }else{ 4690 sqlite3DebugPrintf("\n"); 4691 } 4692 } 4693 } 4694 #endif 4695 4696 /* Swap the roles of aFrom and aTo for the next generation */ 4697 pFrom = aTo; 4698 aTo = aFrom; 4699 aFrom = pFrom; 4700 nFrom = nTo; 4701 } 4702 4703 if( nFrom==0 ){ 4704 sqlite3ErrorMsg(pParse, "no query solution"); 4705 sqlite3DbFreeNN(db, pSpace); 4706 return SQLITE_ERROR; 4707 } 4708 4709 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4710 pFrom = aFrom; 4711 for(ii=1; ii<nFrom; ii++){ 4712 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4713 } 4714 assert( pWInfo->nLevel==nLoop ); 4715 /* Load the lowest cost path into pWInfo */ 4716 for(iLoop=0; iLoop<nLoop; iLoop++){ 4717 WhereLevel *pLevel = pWInfo->a + iLoop; 4718 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4719 pLevel->iFrom = pWLoop->iTab; 4720 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4721 } 4722 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4723 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4724 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4725 && nRowEst 4726 ){ 4727 Bitmask notUsed; 4728 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4729 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4730 if( rc==pWInfo->pResultSet->nExpr ){ 4731 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4732 } 4733 } 4734 pWInfo->bOrderedInnerLoop = 0; 4735 if( pWInfo->pOrderBy ){ 4736 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4737 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4738 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4739 } 4740 }else{ 4741 pWInfo->nOBSat = pFrom->isOrdered; 4742 pWInfo->revMask = pFrom->revLoop; 4743 if( pWInfo->nOBSat<=0 ){ 4744 pWInfo->nOBSat = 0; 4745 if( nLoop>0 ){ 4746 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4747 if( (wsFlags & WHERE_ONEROW)==0 4748 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4749 ){ 4750 Bitmask m = 0; 4751 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4752 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4753 testcase( wsFlags & WHERE_IPK ); 4754 testcase( wsFlags & WHERE_COLUMN_IN ); 4755 if( rc==pWInfo->pOrderBy->nExpr ){ 4756 pWInfo->bOrderedInnerLoop = 1; 4757 pWInfo->revMask = m; 4758 } 4759 } 4760 } 4761 }else if( nLoop 4762 && pWInfo->nOBSat==1 4763 && (pWInfo->wctrlFlags & (WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX))!=0 4764 ){ 4765 pWInfo->bOrderedInnerLoop = 1; 4766 } 4767 } 4768 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4769 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4770 ){ 4771 Bitmask revMask = 0; 4772 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4773 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4774 ); 4775 assert( pWInfo->sorted==0 ); 4776 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4777 pWInfo->sorted = 1; 4778 pWInfo->revMask = revMask; 4779 } 4780 } 4781 } 4782 4783 4784 pWInfo->nRowOut = pFrom->nRow; 4785 4786 /* Free temporary memory and return success */ 4787 sqlite3DbFreeNN(db, pSpace); 4788 return SQLITE_OK; 4789 } 4790 4791 /* 4792 ** Most queries use only a single table (they are not joins) and have 4793 ** simple == constraints against indexed fields. This routine attempts 4794 ** to plan those simple cases using much less ceremony than the 4795 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4796 ** times for the common case. 4797 ** 4798 ** Return non-zero on success, if this query can be handled by this 4799 ** no-frills query planner. Return zero if this query needs the 4800 ** general-purpose query planner. 4801 */ 4802 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4803 WhereInfo *pWInfo; 4804 SrcItem *pItem; 4805 WhereClause *pWC; 4806 WhereTerm *pTerm; 4807 WhereLoop *pLoop; 4808 int iCur; 4809 int j; 4810 Table *pTab; 4811 Index *pIdx; 4812 WhereScan scan; 4813 4814 pWInfo = pBuilder->pWInfo; 4815 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4816 assert( pWInfo->pTabList->nSrc>=1 ); 4817 pItem = pWInfo->pTabList->a; 4818 pTab = pItem->pTab; 4819 if( IsVirtual(pTab) ) return 0; 4820 if( pItem->fg.isIndexedBy ) return 0; 4821 iCur = pItem->iCursor; 4822 pWC = &pWInfo->sWC; 4823 pLoop = pBuilder->pNew; 4824 pLoop->wsFlags = 0; 4825 pLoop->nSkip = 0; 4826 pTerm = whereScanInit(&scan, pWC, iCur, -1, WO_EQ|WO_IS, 0); 4827 while( pTerm && pTerm->prereqRight ) pTerm = whereScanNext(&scan); 4828 if( pTerm ){ 4829 testcase( pTerm->eOperator & WO_IS ); 4830 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4831 pLoop->aLTerm[0] = pTerm; 4832 pLoop->nLTerm = 1; 4833 pLoop->u.btree.nEq = 1; 4834 /* TUNING: Cost of a rowid lookup is 10 */ 4835 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 4836 }else{ 4837 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 4838 int opMask; 4839 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 4840 if( !IsUniqueIndex(pIdx) 4841 || pIdx->pPartIdxWhere!=0 4842 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 4843 ) continue; 4844 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 4845 for(j=0; j<pIdx->nKeyCol; j++){ 4846 pTerm = whereScanInit(&scan, pWC, iCur, j, opMask, pIdx); 4847 while( pTerm && pTerm->prereqRight ) pTerm = whereScanNext(&scan); 4848 if( pTerm==0 ) break; 4849 testcase( pTerm->eOperator & WO_IS ); 4850 pLoop->aLTerm[j] = pTerm; 4851 } 4852 if( j!=pIdx->nKeyCol ) continue; 4853 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 4854 if( pIdx->isCovering || (pItem->colUsed & pIdx->colNotIdxed)==0 ){ 4855 pLoop->wsFlags |= WHERE_IDX_ONLY; 4856 } 4857 pLoop->nLTerm = j; 4858 pLoop->u.btree.nEq = j; 4859 pLoop->u.btree.pIndex = pIdx; 4860 /* TUNING: Cost of a unique index lookup is 15 */ 4861 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 4862 break; 4863 } 4864 } 4865 if( pLoop->wsFlags ){ 4866 pLoop->nOut = (LogEst)1; 4867 pWInfo->a[0].pWLoop = pLoop; 4868 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 4869 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 4870 pWInfo->a[0].iTabCur = iCur; 4871 pWInfo->nRowOut = 1; 4872 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 4873 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 4874 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4875 } 4876 if( scan.iEquiv>1 ) pLoop->wsFlags |= WHERE_TRANSCONS; 4877 #ifdef SQLITE_DEBUG 4878 pLoop->cId = '0'; 4879 #endif 4880 #ifdef WHERETRACE_ENABLED 4881 if( sqlite3WhereTrace ){ 4882 sqlite3DebugPrintf("whereShortCut() used to compute solution\n"); 4883 } 4884 #endif 4885 return 1; 4886 } 4887 return 0; 4888 } 4889 4890 /* 4891 ** Helper function for exprIsDeterministic(). 4892 */ 4893 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 4894 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 4895 pWalker->eCode = 0; 4896 return WRC_Abort; 4897 } 4898 return WRC_Continue; 4899 } 4900 4901 /* 4902 ** Return true if the expression contains no non-deterministic SQL 4903 ** functions. Do not consider non-deterministic SQL functions that are 4904 ** part of sub-select statements. 4905 */ 4906 static int exprIsDeterministic(Expr *p){ 4907 Walker w; 4908 memset(&w, 0, sizeof(w)); 4909 w.eCode = 1; 4910 w.xExprCallback = exprNodeIsDeterministic; 4911 w.xSelectCallback = sqlite3SelectWalkFail; 4912 sqlite3WalkExpr(&w, p); 4913 return w.eCode; 4914 } 4915 4916 4917 #ifdef WHERETRACE_ENABLED 4918 /* 4919 ** Display all WhereLoops in pWInfo 4920 */ 4921 static void showAllWhereLoops(WhereInfo *pWInfo, WhereClause *pWC){ 4922 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4923 WhereLoop *p; 4924 int i; 4925 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4926 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4927 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4928 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 4929 sqlite3WhereLoopPrint(p, pWC); 4930 } 4931 } 4932 } 4933 # define WHERETRACE_ALL_LOOPS(W,C) showAllWhereLoops(W,C) 4934 #else 4935 # define WHERETRACE_ALL_LOOPS(W,C) 4936 #endif 4937 4938 /* Attempt to omit tables from a join that do not affect the result. 4939 ** For a table to not affect the result, the following must be true: 4940 ** 4941 ** 1) The query must not be an aggregate. 4942 ** 2) The table must be the RHS of a LEFT JOIN. 4943 ** 3) Either the query must be DISTINCT, or else the ON or USING clause 4944 ** must contain a constraint that limits the scan of the table to 4945 ** at most a single row. 4946 ** 4) The table must not be referenced by any part of the query apart 4947 ** from its own USING or ON clause. 4948 ** 4949 ** For example, given: 4950 ** 4951 ** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1); 4952 ** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2); 4953 ** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3); 4954 ** 4955 ** then table t2 can be omitted from the following: 4956 ** 4957 ** SELECT v1, v3 FROM t1 4958 ** LEFT JOIN t2 ON (t1.ipk=t2.ipk) 4959 ** LEFT JOIN t3 ON (t1.ipk=t3.ipk) 4960 ** 4961 ** or from: 4962 ** 4963 ** SELECT DISTINCT v1, v3 FROM t1 4964 ** LEFT JOIN t2 4965 ** LEFT JOIN t3 ON (t1.ipk=t3.ipk) 4966 */ 4967 static SQLITE_NOINLINE Bitmask whereOmitNoopJoin( 4968 WhereInfo *pWInfo, 4969 Bitmask notReady 4970 ){ 4971 int i; 4972 Bitmask tabUsed; 4973 4974 /* Preconditions checked by the caller */ 4975 assert( pWInfo->nLevel>=2 ); 4976 assert( OptimizationEnabled(pWInfo->pParse->db, SQLITE_OmitNoopJoin) ); 4977 4978 /* These two preconditions checked by the caller combine to guarantee 4979 ** condition (1) of the header comment */ 4980 assert( pWInfo->pResultSet!=0 ); 4981 assert( 0==(pWInfo->wctrlFlags & WHERE_AGG_DISTINCT) ); 4982 4983 tabUsed = sqlite3WhereExprListUsage(&pWInfo->sMaskSet, pWInfo->pResultSet); 4984 if( pWInfo->pOrderBy ){ 4985 tabUsed |= sqlite3WhereExprListUsage(&pWInfo->sMaskSet, pWInfo->pOrderBy); 4986 } 4987 for(i=pWInfo->nLevel-1; i>=1; i--){ 4988 WhereTerm *pTerm, *pEnd; 4989 SrcItem *pItem; 4990 WhereLoop *pLoop; 4991 pLoop = pWInfo->a[i].pWLoop; 4992 pItem = &pWInfo->pTabList->a[pLoop->iTab]; 4993 if( (pItem->fg.jointype & JT_LEFT)==0 ) continue; 4994 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)==0 4995 && (pLoop->wsFlags & WHERE_ONEROW)==0 4996 ){ 4997 continue; 4998 } 4999 if( (tabUsed & pLoop->maskSelf)!=0 ) continue; 5000 pEnd = pWInfo->sWC.a + pWInfo->sWC.nTerm; 5001 for(pTerm=pWInfo->sWC.a; pTerm<pEnd; pTerm++){ 5002 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 5003 if( !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 5004 || pTerm->pExpr->iRightJoinTable!=pItem->iCursor 5005 ){ 5006 break; 5007 } 5008 } 5009 } 5010 if( pTerm<pEnd ) continue; 5011 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 5012 notReady &= ~pLoop->maskSelf; 5013 for(pTerm=pWInfo->sWC.a; pTerm<pEnd; pTerm++){ 5014 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 5015 pTerm->wtFlags |= TERM_CODED; 5016 } 5017 } 5018 if( i!=pWInfo->nLevel-1 ){ 5019 int nByte = (pWInfo->nLevel-1-i) * sizeof(WhereLevel); 5020 memmove(&pWInfo->a[i], &pWInfo->a[i+1], nByte); 5021 } 5022 pWInfo->nLevel--; 5023 assert( pWInfo->nLevel>0 ); 5024 } 5025 return notReady; 5026 } 5027 5028 /* 5029 ** Check to see if there are any SEARCH loops that might benefit from 5030 ** using a Bloom filter. Consider a Bloom filter if: 5031 ** 5032 ** (1) The SEARCH happens more than N times where N is the number 5033 ** of rows in the table that is being considered for the Bloom 5034 ** filter. 5035 ** (2) Some searches are expected to find zero rows. (This is determined 5036 ** by the WHERE_SELFCULL flag on the term.) 5037 ** (3) Bloom-filter processing is not disabled. (Checked by the 5038 ** caller.) 5039 ** (4) The size of the table being searched is known by ANALYZE. 5040 ** 5041 ** This block of code merely checks to see if a Bloom filter would be 5042 ** appropriate, and if so sets the WHERE_BLOOMFILTER flag on the 5043 ** WhereLoop. The implementation of the Bloom filter comes further 5044 ** down where the code for each WhereLoop is generated. 5045 */ 5046 static SQLITE_NOINLINE void whereCheckIfBloomFilterIsUseful( 5047 const WhereInfo *pWInfo 5048 ){ 5049 int i; 5050 LogEst nSearch; 5051 5052 assert( pWInfo->nLevel>=2 ); 5053 assert( OptimizationEnabled(pWInfo->pParse->db, SQLITE_BloomFilter) ); 5054 nSearch = pWInfo->a[0].pWLoop->nOut; 5055 for(i=1; i<pWInfo->nLevel; i++){ 5056 WhereLoop *pLoop = pWInfo->a[i].pWLoop; 5057 const unsigned int reqFlags = (WHERE_SELFCULL|WHERE_COLUMN_EQ); 5058 if( (pLoop->wsFlags & reqFlags)==reqFlags 5059 /* vvvvvv--- Always the case if WHERE_COLUMN_EQ is defined */ 5060 && ALWAYS((pLoop->wsFlags & (WHERE_IPK|WHERE_INDEXED))!=0) 5061 ){ 5062 SrcItem *pItem = &pWInfo->pTabList->a[pLoop->iTab]; 5063 Table *pTab = pItem->pTab; 5064 pTab->tabFlags |= TF_StatsUsed; 5065 if( nSearch > pTab->nRowLogEst 5066 && (pTab->tabFlags & TF_HasStat1)!=0 5067 ){ 5068 testcase( pItem->fg.jointype & JT_LEFT ); 5069 pLoop->wsFlags |= WHERE_BLOOMFILTER; 5070 pLoop->wsFlags &= ~WHERE_IDX_ONLY; 5071 WHERETRACE(0xffff, ( 5072 "-> use Bloom-filter on loop %c because there are ~%.1e " 5073 "lookups into %s which has only ~%.1e rows\n", 5074 pLoop->cId, (double)sqlite3LogEstToInt(nSearch), pTab->zName, 5075 (double)sqlite3LogEstToInt(pTab->nRowLogEst))); 5076 } 5077 } 5078 nSearch += pLoop->nOut; 5079 } 5080 } 5081 5082 /* 5083 ** Generate the beginning of the loop used for WHERE clause processing. 5084 ** The return value is a pointer to an opaque structure that contains 5085 ** information needed to terminate the loop. Later, the calling routine 5086 ** should invoke sqlite3WhereEnd() with the return value of this function 5087 ** in order to complete the WHERE clause processing. 5088 ** 5089 ** If an error occurs, this routine returns NULL. 5090 ** 5091 ** The basic idea is to do a nested loop, one loop for each table in 5092 ** the FROM clause of a select. (INSERT and UPDATE statements are the 5093 ** same as a SELECT with only a single table in the FROM clause.) For 5094 ** example, if the SQL is this: 5095 ** 5096 ** SELECT * FROM t1, t2, t3 WHERE ...; 5097 ** 5098 ** Then the code generated is conceptually like the following: 5099 ** 5100 ** foreach row1 in t1 do \ Code generated 5101 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 5102 ** foreach row3 in t3 do / 5103 ** ... 5104 ** end \ Code generated 5105 ** end |-- by sqlite3WhereEnd() 5106 ** end / 5107 ** 5108 ** Note that the loops might not be nested in the order in which they 5109 ** appear in the FROM clause if a different order is better able to make 5110 ** use of indices. Note also that when the IN operator appears in 5111 ** the WHERE clause, it might result in additional nested loops for 5112 ** scanning through all values on the right-hand side of the IN. 5113 ** 5114 ** There are Btree cursors associated with each table. t1 uses cursor 5115 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 5116 ** And so forth. This routine generates code to open those VDBE cursors 5117 ** and sqlite3WhereEnd() generates the code to close them. 5118 ** 5119 ** The code that sqlite3WhereBegin() generates leaves the cursors named 5120 ** in pTabList pointing at their appropriate entries. The [...] code 5121 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 5122 ** data from the various tables of the loop. 5123 ** 5124 ** If the WHERE clause is empty, the foreach loops must each scan their 5125 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 5126 ** the tables have indices and there are terms in the WHERE clause that 5127 ** refer to those indices, a complete table scan can be avoided and the 5128 ** code will run much faster. Most of the work of this routine is checking 5129 ** to see if there are indices that can be used to speed up the loop. 5130 ** 5131 ** Terms of the WHERE clause are also used to limit which rows actually 5132 ** make it to the "..." in the middle of the loop. After each "foreach", 5133 ** terms of the WHERE clause that use only terms in that loop and outer 5134 ** loops are evaluated and if false a jump is made around all subsequent 5135 ** inner loops (or around the "..." if the test occurs within the inner- 5136 ** most loop) 5137 ** 5138 ** OUTER JOINS 5139 ** 5140 ** An outer join of tables t1 and t2 is conceptally coded as follows: 5141 ** 5142 ** foreach row1 in t1 do 5143 ** flag = 0 5144 ** foreach row2 in t2 do 5145 ** start: 5146 ** ... 5147 ** flag = 1 5148 ** end 5149 ** if flag==0 then 5150 ** move the row2 cursor to a null row 5151 ** goto start 5152 ** fi 5153 ** end 5154 ** 5155 ** ORDER BY CLAUSE PROCESSING 5156 ** 5157 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 5158 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 5159 ** if there is one. If there is no ORDER BY clause or if this routine 5160 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 5161 ** 5162 ** The iIdxCur parameter is the cursor number of an index. If 5163 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 5164 ** to use for OR clause processing. The WHERE clause should use this 5165 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 5166 ** the first cursor in an array of cursors for all indices. iIdxCur should 5167 ** be used to compute the appropriate cursor depending on which index is 5168 ** used. 5169 */ 5170 WhereInfo *sqlite3WhereBegin( 5171 Parse *pParse, /* The parser context */ 5172 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 5173 Expr *pWhere, /* The WHERE clause */ 5174 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 5175 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 5176 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 5177 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 5178 ** If WHERE_USE_LIMIT, then the limit amount */ 5179 ){ 5180 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 5181 int nTabList; /* Number of elements in pTabList */ 5182 WhereInfo *pWInfo; /* Will become the return value of this function */ 5183 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 5184 Bitmask notReady; /* Cursors that are not yet positioned */ 5185 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 5186 WhereMaskSet *pMaskSet; /* The expression mask set */ 5187 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 5188 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 5189 int ii; /* Loop counter */ 5190 sqlite3 *db; /* Database connection */ 5191 int rc; /* Return code */ 5192 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 5193 5194 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 5195 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 5196 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5197 )); 5198 5199 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 5200 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5201 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 5202 5203 /* Variable initialization */ 5204 db = pParse->db; 5205 memset(&sWLB, 0, sizeof(sWLB)); 5206 5207 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 5208 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 5209 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 5210 sWLB.pOrderBy = pOrderBy; 5211 5212 /* The number of tables in the FROM clause is limited by the number of 5213 ** bits in a Bitmask 5214 */ 5215 testcase( pTabList->nSrc==BMS ); 5216 if( pTabList->nSrc>BMS ){ 5217 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 5218 return 0; 5219 } 5220 5221 /* This function normally generates a nested loop for all tables in 5222 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 5223 ** only generate code for the first table in pTabList and assume that 5224 ** any cursors associated with subsequent tables are uninitialized. 5225 */ 5226 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 5227 5228 /* Allocate and initialize the WhereInfo structure that will become the 5229 ** return value. A single allocation is used to store the WhereInfo 5230 ** struct, the contents of WhereInfo.a[], the WhereClause structure 5231 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 5232 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 5233 ** some architectures. Hence the ROUND8() below. 5234 */ 5235 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 5236 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 5237 if( db->mallocFailed ){ 5238 sqlite3DbFree(db, pWInfo); 5239 pWInfo = 0; 5240 goto whereBeginError; 5241 } 5242 pWInfo->pParse = pParse; 5243 pWInfo->pTabList = pTabList; 5244 pWInfo->pOrderBy = pOrderBy; 5245 pWInfo->pWhere = pWhere; 5246 pWInfo->pResultSet = pResultSet; 5247 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 5248 pWInfo->nLevel = nTabList; 5249 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(pParse); 5250 pWInfo->wctrlFlags = wctrlFlags; 5251 pWInfo->iLimit = iAuxArg; 5252 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 5253 memset(&pWInfo->nOBSat, 0, 5254 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 5255 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 5256 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 5257 pMaskSet = &pWInfo->sMaskSet; 5258 pMaskSet->n = 0; 5259 pMaskSet->ix[0] = -99; /* Initialize ix[0] to a value that can never be 5260 ** a valid cursor number, to avoid an initial 5261 ** test for pMaskSet->n==0 in sqlite3WhereGetMask() */ 5262 sWLB.pWInfo = pWInfo; 5263 sWLB.pWC = &pWInfo->sWC; 5264 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 5265 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 5266 whereLoopInit(sWLB.pNew); 5267 #ifdef SQLITE_DEBUG 5268 sWLB.pNew->cId = '*'; 5269 #endif 5270 5271 /* Split the WHERE clause into separate subexpressions where each 5272 ** subexpression is separated by an AND operator. 5273 */ 5274 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 5275 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 5276 5277 /* Special case: No FROM clause 5278 */ 5279 if( nTabList==0 ){ 5280 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 5281 if( (wctrlFlags & WHERE_WANT_DISTINCT)!=0 5282 && OptimizationEnabled(db, SQLITE_DistinctOpt) 5283 ){ 5284 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 5285 } 5286 ExplainQueryPlan((pParse, 0, "SCAN CONSTANT ROW")); 5287 }else{ 5288 /* Assign a bit from the bitmask to every term in the FROM clause. 5289 ** 5290 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 5291 ** 5292 ** The rule of the previous sentence ensures thta if X is the bitmask for 5293 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 5294 ** Knowing the bitmask for all tables to the left of a left join is 5295 ** important. Ticket #3015. 5296 ** 5297 ** Note that bitmasks are created for all pTabList->nSrc tables in 5298 ** pTabList, not just the first nTabList tables. nTabList is normally 5299 ** equal to pTabList->nSrc but might be shortened to 1 if the 5300 ** WHERE_OR_SUBCLAUSE flag is set. 5301 */ 5302 ii = 0; 5303 do{ 5304 createMask(pMaskSet, pTabList->a[ii].iCursor); 5305 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 5306 }while( (++ii)<pTabList->nSrc ); 5307 #ifdef SQLITE_DEBUG 5308 { 5309 Bitmask mx = 0; 5310 for(ii=0; ii<pTabList->nSrc; ii++){ 5311 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 5312 assert( m>=mx ); 5313 mx = m; 5314 } 5315 } 5316 #endif 5317 } 5318 5319 /* Analyze all of the subexpressions. */ 5320 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 5321 if( db->mallocFailed ) goto whereBeginError; 5322 5323 /* Special case: WHERE terms that do not refer to any tables in the join 5324 ** (constant expressions). Evaluate each such term, and jump over all the 5325 ** generated code if the result is not true. 5326 ** 5327 ** Do not do this if the expression contains non-deterministic functions 5328 ** that are not within a sub-select. This is not strictly required, but 5329 ** preserves SQLite's legacy behaviour in the following two cases: 5330 ** 5331 ** FROM ... WHERE random()>0; -- eval random() once per row 5332 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 5333 */ 5334 for(ii=0; ii<sWLB.pWC->nBase; ii++){ 5335 WhereTerm *pT = &sWLB.pWC->a[ii]; 5336 if( pT->wtFlags & TERM_VIRTUAL ) continue; 5337 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 5338 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 5339 pT->wtFlags |= TERM_CODED; 5340 } 5341 } 5342 5343 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 5344 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 5345 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 5346 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 5347 wctrlFlags &= ~WHERE_WANT_DISTINCT; 5348 pWInfo->wctrlFlags &= ~WHERE_WANT_DISTINCT; 5349 }else if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 5350 /* The DISTINCT marking is pointless. Ignore it. */ 5351 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 5352 }else if( pOrderBy==0 ){ 5353 /* Try to ORDER BY the result set to make distinct processing easier */ 5354 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 5355 pWInfo->pOrderBy = pResultSet; 5356 } 5357 } 5358 5359 /* Construct the WhereLoop objects */ 5360 #if defined(WHERETRACE_ENABLED) 5361 if( sqlite3WhereTrace & 0xffff ){ 5362 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 5363 if( wctrlFlags & WHERE_USE_LIMIT ){ 5364 sqlite3DebugPrintf(", limit: %d", iAuxArg); 5365 } 5366 sqlite3DebugPrintf(")\n"); 5367 if( sqlite3WhereTrace & 0x100 ){ 5368 Select sSelect; 5369 memset(&sSelect, 0, sizeof(sSelect)); 5370 sSelect.selFlags = SF_WhereBegin; 5371 sSelect.pSrc = pTabList; 5372 sSelect.pWhere = pWhere; 5373 sSelect.pOrderBy = pOrderBy; 5374 sSelect.pEList = pResultSet; 5375 sqlite3TreeViewSelect(0, &sSelect, 0); 5376 } 5377 } 5378 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 5379 sqlite3DebugPrintf("---- WHERE clause at start of analysis:\n"); 5380 sqlite3WhereClausePrint(sWLB.pWC); 5381 } 5382 #endif 5383 5384 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 5385 rc = whereLoopAddAll(&sWLB); 5386 if( rc ) goto whereBeginError; 5387 5388 #ifdef SQLITE_ENABLE_STAT4 5389 /* If one or more WhereTerm.truthProb values were used in estimating 5390 ** loop parameters, but then those truthProb values were subsequently 5391 ** changed based on STAT4 information while computing subsequent loops, 5392 ** then we need to rerun the whole loop building process so that all 5393 ** loops will be built using the revised truthProb values. */ 5394 if( sWLB.bldFlags2 & SQLITE_BLDF2_2NDPASS ){ 5395 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC); 5396 WHERETRACE(0xffff, 5397 ("**** Redo all loop computations due to" 5398 " TERM_HIGHTRUTH changes ****\n")); 5399 while( pWInfo->pLoops ){ 5400 WhereLoop *p = pWInfo->pLoops; 5401 pWInfo->pLoops = p->pNextLoop; 5402 whereLoopDelete(db, p); 5403 } 5404 rc = whereLoopAddAll(&sWLB); 5405 if( rc ) goto whereBeginError; 5406 } 5407 #endif 5408 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC); 5409 5410 wherePathSolver(pWInfo, 0); 5411 if( db->mallocFailed ) goto whereBeginError; 5412 if( pWInfo->pOrderBy ){ 5413 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 5414 if( db->mallocFailed ) goto whereBeginError; 5415 } 5416 } 5417 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 5418 pWInfo->revMask = ALLBITS; 5419 } 5420 if( pParse->nErr || db->mallocFailed ){ 5421 goto whereBeginError; 5422 } 5423 #ifdef WHERETRACE_ENABLED 5424 if( sqlite3WhereTrace ){ 5425 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 5426 if( pWInfo->nOBSat>0 ){ 5427 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 5428 } 5429 switch( pWInfo->eDistinct ){ 5430 case WHERE_DISTINCT_UNIQUE: { 5431 sqlite3DebugPrintf(" DISTINCT=unique"); 5432 break; 5433 } 5434 case WHERE_DISTINCT_ORDERED: { 5435 sqlite3DebugPrintf(" DISTINCT=ordered"); 5436 break; 5437 } 5438 case WHERE_DISTINCT_UNORDERED: { 5439 sqlite3DebugPrintf(" DISTINCT=unordered"); 5440 break; 5441 } 5442 } 5443 sqlite3DebugPrintf("\n"); 5444 for(ii=0; ii<pWInfo->nLevel; ii++){ 5445 sqlite3WhereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 5446 } 5447 } 5448 #endif 5449 5450 /* Attempt to omit tables from a join that do not affect the result. 5451 ** See the comment on whereOmitNoopJoin() for further information. 5452 ** 5453 ** This query optimization is factored out into a separate "no-inline" 5454 ** procedure to keep the sqlite3WhereBegin() procedure from becoming 5455 ** too large. If sqlite3WhereBegin() becomes too large, that prevents 5456 ** some C-compiler optimizers from in-lining the 5457 ** sqlite3WhereCodeOneLoopStart() procedure, and it is important to 5458 ** in-line sqlite3WhereCodeOneLoopStart() for performance reasons. 5459 */ 5460 notReady = ~(Bitmask)0; 5461 if( pWInfo->nLevel>=2 5462 && pResultSet!=0 /* these two combine to guarantee */ 5463 && 0==(wctrlFlags & WHERE_AGG_DISTINCT) /* condition (1) above */ 5464 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 5465 ){ 5466 notReady = whereOmitNoopJoin(pWInfo, notReady); 5467 nTabList = pWInfo->nLevel; 5468 assert( nTabList>0 ); 5469 } 5470 5471 /* Check to see if there are any SEARCH loops that might benefit from 5472 ** using a Bloom filter. 5473 */ 5474 if( pWInfo->nLevel>=2 5475 && OptimizationEnabled(db, SQLITE_BloomFilter) 5476 ){ 5477 whereCheckIfBloomFilterIsUseful(pWInfo); 5478 } 5479 5480 #if defined(WHERETRACE_ENABLED) 5481 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 5482 sqlite3DebugPrintf("---- WHERE clause at end of analysis:\n"); 5483 sqlite3WhereClausePrint(sWLB.pWC); 5484 } 5485 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 5486 #endif 5487 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 5488 5489 /* If the caller is an UPDATE or DELETE statement that is requesting 5490 ** to use a one-pass algorithm, determine if this is appropriate. 5491 ** 5492 ** A one-pass approach can be used if the caller has requested one 5493 ** and either (a) the scan visits at most one row or (b) each 5494 ** of the following are true: 5495 ** 5496 ** * the caller has indicated that a one-pass approach can be used 5497 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and 5498 ** * the table is not a virtual table, and 5499 ** * either the scan does not use the OR optimization or the caller 5500 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified 5501 ** for DELETE). 5502 ** 5503 ** The last qualification is because an UPDATE statement uses 5504 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can 5505 ** use a one-pass approach, and this is not set accurately for scans 5506 ** that use the OR optimization. 5507 */ 5508 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 5509 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 5510 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 5511 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 5512 assert( !(wsFlags & WHERE_VIRTUALTABLE) || IsVirtual(pTabList->a[0].pTab) ); 5513 if( bOnerow || ( 5514 0!=(wctrlFlags & WHERE_ONEPASS_MULTIROW) 5515 && !IsVirtual(pTabList->a[0].pTab) 5516 && (0==(wsFlags & WHERE_MULTI_OR) || (wctrlFlags & WHERE_DUPLICATES_OK)) 5517 )){ 5518 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 5519 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 5520 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 5521 bFordelete = OPFLAG_FORDELETE; 5522 } 5523 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 5524 } 5525 } 5526 } 5527 5528 /* Open all tables in the pTabList and any indices selected for 5529 ** searching those tables. 5530 */ 5531 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 5532 Table *pTab; /* Table to open */ 5533 int iDb; /* Index of database containing table/index */ 5534 SrcItem *pTabItem; 5535 5536 pTabItem = &pTabList->a[pLevel->iFrom]; 5537 pTab = pTabItem->pTab; 5538 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 5539 pLoop = pLevel->pWLoop; 5540 if( (pTab->tabFlags & TF_Ephemeral)!=0 || IsView(pTab) ){ 5541 /* Do nothing */ 5542 }else 5543 #ifndef SQLITE_OMIT_VIRTUALTABLE 5544 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 5545 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 5546 int iCur = pTabItem->iCursor; 5547 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 5548 }else if( IsVirtual(pTab) ){ 5549 /* noop */ 5550 }else 5551 #endif 5552 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 5553 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 5554 int op = OP_OpenRead; 5555 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5556 op = OP_OpenWrite; 5557 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 5558 }; 5559 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 5560 assert( pTabItem->iCursor==pLevel->iTabCur ); 5561 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 5562 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 5563 if( pWInfo->eOnePass==ONEPASS_OFF 5564 && pTab->nCol<BMS 5565 && (pTab->tabFlags & (TF_HasGenerated|TF_WithoutRowid))==0 5566 ){ 5567 /* If we know that only a prefix of the record will be used, 5568 ** it is advantageous to reduce the "column count" field in 5569 ** the P4 operand of the OP_OpenRead/Write opcode. */ 5570 Bitmask b = pTabItem->colUsed; 5571 int n = 0; 5572 for(; b; b=b>>1, n++){} 5573 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 5574 assert( n<=pTab->nCol ); 5575 } 5576 #ifdef SQLITE_ENABLE_CURSOR_HINTS 5577 if( pLoop->u.btree.pIndex!=0 ){ 5578 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 5579 }else 5580 #endif 5581 { 5582 sqlite3VdbeChangeP5(v, bFordelete); 5583 } 5584 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5585 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 5586 (const u8*)&pTabItem->colUsed, P4_INT64); 5587 #endif 5588 }else{ 5589 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 5590 } 5591 if( pLoop->wsFlags & WHERE_INDEXED ){ 5592 Index *pIx = pLoop->u.btree.pIndex; 5593 int iIndexCur; 5594 int op = OP_OpenRead; 5595 /* iAuxArg is always set to a positive value if ONEPASS is possible */ 5596 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 5597 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 5598 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 5599 ){ 5600 /* This is one term of an OR-optimization using the PRIMARY KEY of a 5601 ** WITHOUT ROWID table. No need for a separate index */ 5602 iIndexCur = pLevel->iTabCur; 5603 op = 0; 5604 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5605 Index *pJ = pTabItem->pTab->pIndex; 5606 iIndexCur = iAuxArg; 5607 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 5608 while( ALWAYS(pJ) && pJ!=pIx ){ 5609 iIndexCur++; 5610 pJ = pJ->pNext; 5611 } 5612 op = OP_OpenWrite; 5613 pWInfo->aiCurOnePass[1] = iIndexCur; 5614 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 5615 iIndexCur = iAuxArg; 5616 op = OP_ReopenIdx; 5617 }else{ 5618 iIndexCur = pParse->nTab++; 5619 } 5620 pLevel->iIdxCur = iIndexCur; 5621 assert( pIx->pSchema==pTab->pSchema ); 5622 assert( iIndexCur>=0 ); 5623 if( op ){ 5624 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 5625 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 5626 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 5627 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 5628 && (pLoop->wsFlags & WHERE_BIGNULL_SORT)==0 5629 && (pLoop->wsFlags & WHERE_IN_SEEKSCAN)==0 5630 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 5631 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 5632 ){ 5633 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); 5634 } 5635 VdbeComment((v, "%s", pIx->zName)); 5636 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5637 { 5638 u64 colUsed = 0; 5639 int ii, jj; 5640 for(ii=0; ii<pIx->nColumn; ii++){ 5641 jj = pIx->aiColumn[ii]; 5642 if( jj<0 ) continue; 5643 if( jj>63 ) jj = 63; 5644 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 5645 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 5646 } 5647 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 5648 (u8*)&colUsed, P4_INT64); 5649 } 5650 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 5651 } 5652 } 5653 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 5654 } 5655 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 5656 if( db->mallocFailed ) goto whereBeginError; 5657 5658 /* Generate the code to do the search. Each iteration of the for 5659 ** loop below generates code for a single nested loop of the VM 5660 ** program. 5661 */ 5662 for(ii=0; ii<nTabList; ii++){ 5663 int addrExplain; 5664 int wsFlags; 5665 if( pParse->nErr ) goto whereBeginError; 5666 pLevel = &pWInfo->a[ii]; 5667 wsFlags = pLevel->pWLoop->wsFlags; 5668 if( (wsFlags & (WHERE_AUTO_INDEX|WHERE_BLOOMFILTER))!=0 ){ 5669 if( (wsFlags & WHERE_AUTO_INDEX)!=0 ){ 5670 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 5671 constructAutomaticIndex(pParse, &pWInfo->sWC, 5672 &pTabList->a[pLevel->iFrom], notReady, pLevel); 5673 #endif 5674 }else{ 5675 sqlite3ConstructBloomFilter(pWInfo, ii, pLevel, notReady); 5676 } 5677 if( db->mallocFailed ) goto whereBeginError; 5678 } 5679 addrExplain = sqlite3WhereExplainOneScan( 5680 pParse, pTabList, pLevel, wctrlFlags 5681 ); 5682 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 5683 notReady = sqlite3WhereCodeOneLoopStart(pParse,v,pWInfo,ii,pLevel,notReady); 5684 pWInfo->iContinue = pLevel->addrCont; 5685 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 5686 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 5687 } 5688 } 5689 5690 /* Done. */ 5691 VdbeModuleComment((v, "Begin WHERE-core")); 5692 pWInfo->iEndWhere = sqlite3VdbeCurrentAddr(v); 5693 return pWInfo; 5694 5695 /* Jump here if malloc fails */ 5696 whereBeginError: 5697 if( pWInfo ){ 5698 testcase( pWInfo->pExprMods!=0 ); 5699 whereUndoExprMods(pWInfo); 5700 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5701 whereInfoFree(db, pWInfo); 5702 } 5703 return 0; 5704 } 5705 5706 /* 5707 ** Part of sqlite3WhereEnd() will rewrite opcodes to reference the 5708 ** index rather than the main table. In SQLITE_DEBUG mode, we want 5709 ** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine 5710 ** does that. 5711 */ 5712 #ifndef SQLITE_DEBUG 5713 # define OpcodeRewriteTrace(D,K,P) /* no-op */ 5714 #else 5715 # define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P) 5716 static void sqlite3WhereOpcodeRewriteTrace( 5717 sqlite3 *db, 5718 int pc, 5719 VdbeOp *pOp 5720 ){ 5721 if( (db->flags & SQLITE_VdbeAddopTrace)==0 ) return; 5722 sqlite3VdbePrintOp(0, pc, pOp); 5723 } 5724 #endif 5725 5726 /* 5727 ** Generate the end of the WHERE loop. See comments on 5728 ** sqlite3WhereBegin() for additional information. 5729 */ 5730 void sqlite3WhereEnd(WhereInfo *pWInfo){ 5731 Parse *pParse = pWInfo->pParse; 5732 Vdbe *v = pParse->pVdbe; 5733 int i; 5734 WhereLevel *pLevel; 5735 WhereLoop *pLoop; 5736 SrcList *pTabList = pWInfo->pTabList; 5737 sqlite3 *db = pParse->db; 5738 int iEnd = sqlite3VdbeCurrentAddr(v); 5739 5740 /* Generate loop termination code. 5741 */ 5742 VdbeModuleComment((v, "End WHERE-core")); 5743 for(i=pWInfo->nLevel-1; i>=0; i--){ 5744 int addr; 5745 pLevel = &pWInfo->a[i]; 5746 pLoop = pLevel->pWLoop; 5747 if( pLevel->op!=OP_Noop ){ 5748 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5749 int addrSeek = 0; 5750 Index *pIdx; 5751 int n; 5752 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 5753 && i==pWInfo->nLevel-1 /* Ticket [ef9318757b152e3] 2017-10-21 */ 5754 && (pLoop->wsFlags & WHERE_INDEXED)!=0 5755 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 5756 && (n = pLoop->u.btree.nDistinctCol)>0 5757 && pIdx->aiRowLogEst[n]>=36 5758 ){ 5759 int r1 = pParse->nMem+1; 5760 int j, op; 5761 for(j=0; j<n; j++){ 5762 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 5763 } 5764 pParse->nMem += n+1; 5765 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 5766 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 5767 VdbeCoverageIf(v, op==OP_SeekLT); 5768 VdbeCoverageIf(v, op==OP_SeekGT); 5769 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 5770 } 5771 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 5772 /* The common case: Advance to the next row */ 5773 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5774 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 5775 sqlite3VdbeChangeP5(v, pLevel->p5); 5776 VdbeCoverage(v); 5777 VdbeCoverageIf(v, pLevel->op==OP_Next); 5778 VdbeCoverageIf(v, pLevel->op==OP_Prev); 5779 VdbeCoverageIf(v, pLevel->op==OP_VNext); 5780 if( pLevel->regBignull ){ 5781 sqlite3VdbeResolveLabel(v, pLevel->addrBignull); 5782 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, pLevel->regBignull, pLevel->p2-1); 5783 VdbeCoverage(v); 5784 } 5785 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5786 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 5787 #endif 5788 }else{ 5789 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5790 } 5791 if( (pLoop->wsFlags & WHERE_IN_ABLE)!=0 && pLevel->u.in.nIn>0 ){ 5792 struct InLoop *pIn; 5793 int j; 5794 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 5795 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 5796 assert( sqlite3VdbeGetOp(v, pIn->addrInTop+1)->opcode==OP_IsNull 5797 || pParse->db->mallocFailed ); 5798 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5799 if( pIn->eEndLoopOp!=OP_Noop ){ 5800 if( pIn->nPrefix ){ 5801 int bEarlyOut = 5802 (pLoop->wsFlags & WHERE_VIRTUALTABLE)==0 5803 && (pLoop->wsFlags & WHERE_IN_EARLYOUT)!=0; 5804 if( pLevel->iLeftJoin ){ 5805 /* For LEFT JOIN queries, cursor pIn->iCur may not have been 5806 ** opened yet. This occurs for WHERE clauses such as 5807 ** "a = ? AND b IN (...)", where the index is on (a, b). If 5808 ** the RHS of the (a=?) is NULL, then the "b IN (...)" may 5809 ** never have been coded, but the body of the loop run to 5810 ** return the null-row. So, if the cursor is not open yet, 5811 ** jump over the OP_Next or OP_Prev instruction about to 5812 ** be coded. */ 5813 sqlite3VdbeAddOp2(v, OP_IfNotOpen, pIn->iCur, 5814 sqlite3VdbeCurrentAddr(v) + 2 + bEarlyOut); 5815 VdbeCoverage(v); 5816 } 5817 if( bEarlyOut ){ 5818 sqlite3VdbeAddOp4Int(v, OP_IfNoHope, pLevel->iIdxCur, 5819 sqlite3VdbeCurrentAddr(v)+2, 5820 pIn->iBase, pIn->nPrefix); 5821 VdbeCoverage(v); 5822 /* Retarget the OP_IsNull against the left operand of IN so 5823 ** it jumps past the OP_IfNoHope. This is because the 5824 ** OP_IsNull also bypasses the OP_Affinity opcode that is 5825 ** required by OP_IfNoHope. */ 5826 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5827 } 5828 } 5829 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 5830 VdbeCoverage(v); 5831 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Prev); 5832 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Next); 5833 } 5834 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 5835 } 5836 } 5837 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 5838 if( pLevel->addrSkip ){ 5839 sqlite3VdbeGoto(v, pLevel->addrSkip); 5840 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 5841 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 5842 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 5843 } 5844 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 5845 if( pLevel->addrLikeRep ){ 5846 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 5847 pLevel->addrLikeRep); 5848 VdbeCoverage(v); 5849 } 5850 #endif 5851 if( pLevel->iLeftJoin ){ 5852 int ws = pLoop->wsFlags; 5853 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 5854 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 5855 if( (ws & WHERE_IDX_ONLY)==0 ){ 5856 assert( pLevel->iTabCur==pTabList->a[pLevel->iFrom].iCursor ); 5857 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iTabCur); 5858 } 5859 if( (ws & WHERE_INDEXED) 5860 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCoveringIdx) 5861 ){ 5862 if( ws & WHERE_MULTI_OR ){ 5863 Index *pIx = pLevel->u.pCoveringIdx; 5864 int iDb = sqlite3SchemaToIndex(db, pIx->pSchema); 5865 sqlite3VdbeAddOp3(v, OP_ReopenIdx, pLevel->iIdxCur, pIx->tnum, iDb); 5866 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 5867 } 5868 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 5869 } 5870 if( pLevel->op==OP_Return ){ 5871 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 5872 }else{ 5873 sqlite3VdbeGoto(v, pLevel->addrFirst); 5874 } 5875 sqlite3VdbeJumpHere(v, addr); 5876 } 5877 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 5878 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 5879 } 5880 5881 /* The "break" point is here, just past the end of the outer loop. 5882 ** Set it. 5883 */ 5884 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5885 5886 assert( pWInfo->nLevel<=pTabList->nSrc ); 5887 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5888 int k, last; 5889 VdbeOp *pOp, *pLastOp; 5890 Index *pIdx = 0; 5891 SrcItem *pTabItem = &pTabList->a[pLevel->iFrom]; 5892 Table *pTab = pTabItem->pTab; 5893 assert( pTab!=0 ); 5894 pLoop = pLevel->pWLoop; 5895 5896 /* For a co-routine, change all OP_Column references to the table of 5897 ** the co-routine into OP_Copy of result contained in a register. 5898 ** OP_Rowid becomes OP_Null. 5899 */ 5900 if( pTabItem->fg.viaCoroutine ){ 5901 testcase( pParse->db->mallocFailed ); 5902 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 5903 pTabItem->regResult, 0); 5904 continue; 5905 } 5906 5907 #ifdef SQLITE_ENABLE_EARLY_CURSOR_CLOSE 5908 /* Close all of the cursors that were opened by sqlite3WhereBegin. 5909 ** Except, do not close cursors that will be reused by the OR optimization 5910 ** (WHERE_OR_SUBCLAUSE). And do not close the OP_OpenWrite cursors 5911 ** created for the ONEPASS optimization. 5912 */ 5913 if( (pTab->tabFlags & TF_Ephemeral)==0 5914 && !IsView(pTab) 5915 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5916 ){ 5917 int ws = pLoop->wsFlags; 5918 if( pWInfo->eOnePass==ONEPASS_OFF && (ws & WHERE_IDX_ONLY)==0 ){ 5919 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); 5920 } 5921 if( (ws & WHERE_INDEXED)!=0 5922 && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0 5923 && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1] 5924 ){ 5925 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); 5926 } 5927 } 5928 #endif 5929 5930 /* If this scan uses an index, make VDBE code substitutions to read data 5931 ** from the index instead of from the table where possible. In some cases 5932 ** this optimization prevents the table from ever being read, which can 5933 ** yield a significant performance boost. 5934 ** 5935 ** Calls to the code generator in between sqlite3WhereBegin and 5936 ** sqlite3WhereEnd will have created code that references the table 5937 ** directly. This loop scans all that code looking for opcodes 5938 ** that reference the table and converts them into opcodes that 5939 ** reference the index. 5940 */ 5941 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 5942 pIdx = pLoop->u.btree.pIndex; 5943 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 5944 pIdx = pLevel->u.pCoveringIdx; 5945 } 5946 if( pIdx 5947 && !db->mallocFailed 5948 ){ 5949 if( pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable) ){ 5950 last = iEnd; 5951 }else{ 5952 last = pWInfo->iEndWhere; 5953 } 5954 k = pLevel->addrBody + 1; 5955 #ifdef SQLITE_DEBUG 5956 if( db->flags & SQLITE_VdbeAddopTrace ){ 5957 printf("TRANSLATE opcodes in range %d..%d\n", k, last-1); 5958 } 5959 /* Proof that the "+1" on the k value above is safe */ 5960 pOp = sqlite3VdbeGetOp(v, k - 1); 5961 assert( pOp->opcode!=OP_Column || pOp->p1!=pLevel->iTabCur ); 5962 assert( pOp->opcode!=OP_Rowid || pOp->p1!=pLevel->iTabCur ); 5963 assert( pOp->opcode!=OP_IfNullRow || pOp->p1!=pLevel->iTabCur ); 5964 #endif 5965 pOp = sqlite3VdbeGetOp(v, k); 5966 pLastOp = pOp + (last - k); 5967 assert( pOp<=pLastOp ); 5968 do{ 5969 if( pOp->p1!=pLevel->iTabCur ){ 5970 /* no-op */ 5971 }else if( pOp->opcode==OP_Column 5972 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC 5973 || pOp->opcode==OP_Offset 5974 #endif 5975 ){ 5976 int x = pOp->p2; 5977 assert( pIdx->pTable==pTab ); 5978 if( !HasRowid(pTab) ){ 5979 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 5980 x = pPk->aiColumn[x]; 5981 assert( x>=0 ); 5982 }else{ 5983 testcase( x!=sqlite3StorageColumnToTable(pTab,x) ); 5984 x = sqlite3StorageColumnToTable(pTab,x); 5985 } 5986 x = sqlite3TableColumnToIndex(pIdx, x); 5987 if( x>=0 ){ 5988 pOp->p2 = x; 5989 pOp->p1 = pLevel->iIdxCur; 5990 OpcodeRewriteTrace(db, k, pOp); 5991 } 5992 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 5993 || pWInfo->eOnePass ); 5994 }else if( pOp->opcode==OP_Rowid ){ 5995 pOp->p1 = pLevel->iIdxCur; 5996 pOp->opcode = OP_IdxRowid; 5997 OpcodeRewriteTrace(db, k, pOp); 5998 }else if( pOp->opcode==OP_IfNullRow ){ 5999 pOp->p1 = pLevel->iIdxCur; 6000 OpcodeRewriteTrace(db, k, pOp); 6001 } 6002 #ifdef SQLITE_DEBUG 6003 k++; 6004 #endif 6005 }while( (++pOp)<pLastOp ); 6006 #ifdef SQLITE_DEBUG 6007 if( db->flags & SQLITE_VdbeAddopTrace ) printf("TRANSLATE complete\n"); 6008 #endif 6009 } 6010 } 6011 6012 /* Final cleanup 6013 */ 6014 if( pWInfo->pExprMods ) whereUndoExprMods(pWInfo); 6015 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 6016 whereInfoFree(db, pWInfo); 6017 return; 6018 } 6019