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 int eDistinct; /* Value to return from sqlite3_vtab_distinct() */ 36 u32 mIn; /* Mask of terms that are <col> IN (...) */ 37 u32 mHandleIn; /* Terms that vtab will handle as <col> IN (...) */ 38 sqlite3_value *aRhs[1]; /* RHS values for constraints. MUST BE LAST 39 ** because extra space is allocated to hold up 40 ** to nTerm such values */ 41 }; 42 43 /* Forward declaration of methods */ 44 static int whereLoopResize(sqlite3*, WhereLoop*, int); 45 46 /* 47 ** Return the estimated number of output rows from a WHERE clause 48 */ 49 LogEst sqlite3WhereOutputRowCount(WhereInfo *pWInfo){ 50 return pWInfo->nRowOut; 51 } 52 53 /* 54 ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this 55 ** WHERE clause returns outputs for DISTINCT processing. 56 */ 57 int sqlite3WhereIsDistinct(WhereInfo *pWInfo){ 58 return pWInfo->eDistinct; 59 } 60 61 /* 62 ** Return the number of ORDER BY terms that are satisfied by the 63 ** WHERE clause. A return of 0 means that the output must be 64 ** completely sorted. A return equal to the number of ORDER BY 65 ** terms means that no sorting is needed at all. A return that 66 ** is positive but less than the number of ORDER BY terms means that 67 ** block sorting is required. 68 */ 69 int sqlite3WhereIsOrdered(WhereInfo *pWInfo){ 70 return pWInfo->nOBSat; 71 } 72 73 /* 74 ** In the ORDER BY LIMIT optimization, if the inner-most loop is known 75 ** to emit rows in increasing order, and if the last row emitted by the 76 ** inner-most loop did not fit within the sorter, then we can skip all 77 ** subsequent rows for the current iteration of the inner loop (because they 78 ** will not fit in the sorter either) and continue with the second inner 79 ** loop - the loop immediately outside the inner-most. 80 ** 81 ** When a row does not fit in the sorter (because the sorter already 82 ** holds LIMIT+OFFSET rows that are smaller), then a jump is made to the 83 ** label returned by this function. 84 ** 85 ** If the ORDER BY LIMIT optimization applies, the jump destination should 86 ** be the continuation for the second-inner-most loop. If the ORDER BY 87 ** LIMIT optimization does not apply, then the jump destination should 88 ** be the continuation for the inner-most loop. 89 ** 90 ** It is always safe for this routine to return the continuation of the 91 ** inner-most loop, in the sense that a correct answer will result. 92 ** Returning the continuation the second inner loop is an optimization 93 ** that might make the code run a little faster, but should not change 94 ** the final answer. 95 */ 96 int sqlite3WhereOrderByLimitOptLabel(WhereInfo *pWInfo){ 97 WhereLevel *pInner; 98 if( !pWInfo->bOrderedInnerLoop ){ 99 /* The ORDER BY LIMIT optimization does not apply. Jump to the 100 ** continuation of the inner-most loop. */ 101 return pWInfo->iContinue; 102 } 103 pInner = &pWInfo->a[pWInfo->nLevel-1]; 104 assert( pInner->addrNxt!=0 ); 105 return pInner->addrNxt; 106 } 107 108 /* 109 ** While generating code for the min/max optimization, after handling 110 ** the aggregate-step call to min() or max(), check to see if any 111 ** additional looping is required. If the output order is such that 112 ** we are certain that the correct answer has already been found, then 113 ** code an OP_Goto to by pass subsequent processing. 114 ** 115 ** Any extra OP_Goto that is coded here is an optimization. The 116 ** correct answer should be obtained regardless. This OP_Goto just 117 ** makes the answer appear faster. 118 */ 119 void sqlite3WhereMinMaxOptEarlyOut(Vdbe *v, WhereInfo *pWInfo){ 120 WhereLevel *pInner; 121 int i; 122 if( !pWInfo->bOrderedInnerLoop ) return; 123 if( pWInfo->nOBSat==0 ) return; 124 for(i=pWInfo->nLevel-1; i>=0; i--){ 125 pInner = &pWInfo->a[i]; 126 if( (pInner->pWLoop->wsFlags & WHERE_COLUMN_IN)!=0 ){ 127 sqlite3VdbeGoto(v, pInner->addrNxt); 128 return; 129 } 130 } 131 sqlite3VdbeGoto(v, pWInfo->iBreak); 132 } 133 134 /* 135 ** Return the VDBE address or label to jump to in order to continue 136 ** immediately with the next row of a WHERE clause. 137 */ 138 int sqlite3WhereContinueLabel(WhereInfo *pWInfo){ 139 assert( pWInfo->iContinue!=0 ); 140 return pWInfo->iContinue; 141 } 142 143 /* 144 ** Return the VDBE address or label to jump to in order to break 145 ** out of a WHERE loop. 146 */ 147 int sqlite3WhereBreakLabel(WhereInfo *pWInfo){ 148 return pWInfo->iBreak; 149 } 150 151 /* 152 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to 153 ** operate directly on the rowids returned by a WHERE clause. Return 154 ** ONEPASS_SINGLE (1) if the statement can operation directly because only 155 ** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass 156 ** optimization can be used on multiple 157 ** 158 ** If the ONEPASS optimization is used (if this routine returns true) 159 ** then also write the indices of open cursors used by ONEPASS 160 ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data 161 ** table and iaCur[1] gets the cursor used by an auxiliary index. 162 ** Either value may be -1, indicating that cursor is not used. 163 ** Any cursors returned will have been opened for writing. 164 ** 165 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is 166 ** unable to use the ONEPASS optimization. 167 */ 168 int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){ 169 memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2); 170 #ifdef WHERETRACE_ENABLED 171 if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){ 172 sqlite3DebugPrintf("%s cursors: %d %d\n", 173 pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI", 174 aiCur[0], aiCur[1]); 175 } 176 #endif 177 return pWInfo->eOnePass; 178 } 179 180 /* 181 ** Return TRUE if the WHERE loop uses the OP_DeferredSeek opcode to move 182 ** the data cursor to the row selected by the index cursor. 183 */ 184 int sqlite3WhereUsesDeferredSeek(WhereInfo *pWInfo){ 185 return pWInfo->bDeferredSeek; 186 } 187 188 /* 189 ** Move the content of pSrc into pDest 190 */ 191 static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){ 192 pDest->n = pSrc->n; 193 memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0])); 194 } 195 196 /* 197 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet. 198 ** 199 ** The new entry might overwrite an existing entry, or it might be 200 ** appended, or it might be discarded. Do whatever is the right thing 201 ** so that pSet keeps the N_OR_COST best entries seen so far. 202 */ 203 static int whereOrInsert( 204 WhereOrSet *pSet, /* The WhereOrSet to be updated */ 205 Bitmask prereq, /* Prerequisites of the new entry */ 206 LogEst rRun, /* Run-cost of the new entry */ 207 LogEst nOut /* Number of outputs for the new entry */ 208 ){ 209 u16 i; 210 WhereOrCost *p; 211 for(i=pSet->n, p=pSet->a; i>0; i--, p++){ 212 if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){ 213 goto whereOrInsert_done; 214 } 215 if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){ 216 return 0; 217 } 218 } 219 if( pSet->n<N_OR_COST ){ 220 p = &pSet->a[pSet->n++]; 221 p->nOut = nOut; 222 }else{ 223 p = pSet->a; 224 for(i=1; i<pSet->n; i++){ 225 if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i; 226 } 227 if( p->rRun<=rRun ) return 0; 228 } 229 whereOrInsert_done: 230 p->prereq = prereq; 231 p->rRun = rRun; 232 if( p->nOut>nOut ) p->nOut = nOut; 233 return 1; 234 } 235 236 /* 237 ** Return the bitmask for the given cursor number. Return 0 if 238 ** iCursor is not in the set. 239 */ 240 Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){ 241 int i; 242 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 ); 243 assert( pMaskSet->n>0 || pMaskSet->ix[0]<0 ); 244 assert( iCursor>=-1 ); 245 if( pMaskSet->ix[0]==iCursor ){ 246 return 1; 247 } 248 for(i=1; i<pMaskSet->n; i++){ 249 if( pMaskSet->ix[i]==iCursor ){ 250 return MASKBIT(i); 251 } 252 } 253 return 0; 254 } 255 256 /* 257 ** Create a new mask for cursor iCursor. 258 ** 259 ** There is one cursor per table in the FROM clause. The number of 260 ** tables in the FROM clause is limited by a test early in the 261 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] 262 ** array will never overflow. 263 */ 264 static void createMask(WhereMaskSet *pMaskSet, int iCursor){ 265 assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); 266 pMaskSet->ix[pMaskSet->n++] = iCursor; 267 } 268 269 /* 270 ** If the right-hand branch of the expression is a TK_COLUMN, then return 271 ** a pointer to the right-hand branch. Otherwise, return NULL. 272 */ 273 static Expr *whereRightSubexprIsColumn(Expr *p){ 274 p = sqlite3ExprSkipCollateAndLikely(p->pRight); 275 if( ALWAYS(p!=0) && p->op==TK_COLUMN && !ExprHasProperty(p, EP_FixedCol) ){ 276 return p; 277 } 278 return 0; 279 } 280 281 /* 282 ** Advance to the next WhereTerm that matches according to the criteria 283 ** established when the pScan object was initialized by whereScanInit(). 284 ** Return NULL if there are no more matching WhereTerms. 285 */ 286 static WhereTerm *whereScanNext(WhereScan *pScan){ 287 int iCur; /* The cursor on the LHS of the term */ 288 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */ 289 Expr *pX; /* An expression being tested */ 290 WhereClause *pWC; /* Shorthand for pScan->pWC */ 291 WhereTerm *pTerm; /* The term being tested */ 292 int k = pScan->k; /* Where to start scanning */ 293 294 assert( pScan->iEquiv<=pScan->nEquiv ); 295 pWC = pScan->pWC; 296 while(1){ 297 iColumn = pScan->aiColumn[pScan->iEquiv-1]; 298 iCur = pScan->aiCur[pScan->iEquiv-1]; 299 assert( pWC!=0 ); 300 assert( iCur>=0 ); 301 do{ 302 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){ 303 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 || pTerm->leftCursor<0 ); 304 if( pTerm->leftCursor==iCur 305 && pTerm->u.x.leftColumn==iColumn 306 && (iColumn!=XN_EXPR 307 || sqlite3ExprCompareSkip(pTerm->pExpr->pLeft, 308 pScan->pIdxExpr,iCur)==0) 309 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin)) 310 ){ 311 if( (pTerm->eOperator & WO_EQUIV)!=0 312 && pScan->nEquiv<ArraySize(pScan->aiCur) 313 && (pX = whereRightSubexprIsColumn(pTerm->pExpr))!=0 314 ){ 315 int j; 316 for(j=0; j<pScan->nEquiv; j++){ 317 if( pScan->aiCur[j]==pX->iTable 318 && pScan->aiColumn[j]==pX->iColumn ){ 319 break; 320 } 321 } 322 if( j==pScan->nEquiv ){ 323 pScan->aiCur[j] = pX->iTable; 324 pScan->aiColumn[j] = pX->iColumn; 325 pScan->nEquiv++; 326 } 327 } 328 if( (pTerm->eOperator & pScan->opMask)!=0 ){ 329 /* Verify the affinity and collating sequence match */ 330 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){ 331 CollSeq *pColl; 332 Parse *pParse = pWC->pWInfo->pParse; 333 pX = pTerm->pExpr; 334 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){ 335 continue; 336 } 337 assert(pX->pLeft); 338 pColl = sqlite3ExprCompareCollSeq(pParse, pX); 339 if( pColl==0 ) pColl = pParse->db->pDfltColl; 340 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){ 341 continue; 342 } 343 } 344 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0 345 && (pX = pTerm->pExpr->pRight, ALWAYS(pX!=0)) 346 && pX->op==TK_COLUMN 347 && pX->iTable==pScan->aiCur[0] 348 && pX->iColumn==pScan->aiColumn[0] 349 ){ 350 testcase( pTerm->eOperator & WO_IS ); 351 continue; 352 } 353 pScan->pWC = pWC; 354 pScan->k = k+1; 355 #ifdef WHERETRACE_ENABLED 356 if( sqlite3WhereTrace & 0x20000 ){ 357 int ii; 358 sqlite3DebugPrintf("SCAN-TERM %p: nEquiv=%d", 359 pTerm, pScan->nEquiv); 360 for(ii=0; ii<pScan->nEquiv; ii++){ 361 sqlite3DebugPrintf(" {%d:%d}", 362 pScan->aiCur[ii], pScan->aiColumn[ii]); 363 } 364 sqlite3DebugPrintf("\n"); 365 } 366 #endif 367 return pTerm; 368 } 369 } 370 } 371 pWC = pWC->pOuter; 372 k = 0; 373 }while( pWC!=0 ); 374 if( pScan->iEquiv>=pScan->nEquiv ) break; 375 pWC = pScan->pOrigWC; 376 k = 0; 377 pScan->iEquiv++; 378 } 379 return 0; 380 } 381 382 /* 383 ** This is whereScanInit() for the case of an index on an expression. 384 ** It is factored out into a separate tail-recursion subroutine so that 385 ** the normal whereScanInit() routine, which is a high-runner, does not 386 ** need to push registers onto the stack as part of its prologue. 387 */ 388 static SQLITE_NOINLINE WhereTerm *whereScanInitIndexExpr(WhereScan *pScan){ 389 pScan->idxaff = sqlite3ExprAffinity(pScan->pIdxExpr); 390 return whereScanNext(pScan); 391 } 392 393 /* 394 ** Initialize a WHERE clause scanner object. Return a pointer to the 395 ** first match. Return NULL if there are no matches. 396 ** 397 ** The scanner will be searching the WHERE clause pWC. It will look 398 ** for terms of the form "X <op> <expr>" where X is column iColumn of table 399 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx 400 ** must be one of the indexes of table iCur. 401 ** 402 ** The <op> must be one of the operators described by opMask. 403 ** 404 ** If the search is for X and the WHERE clause contains terms of the 405 ** form X=Y then this routine might also return terms of the form 406 ** "Y <op> <expr>". The number of levels of transitivity is limited, 407 ** but is enough to handle most commonly occurring SQL statements. 408 ** 409 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with 410 ** index pIdx. 411 */ 412 static WhereTerm *whereScanInit( 413 WhereScan *pScan, /* The WhereScan object being initialized */ 414 WhereClause *pWC, /* The WHERE clause to be scanned */ 415 int iCur, /* Cursor to scan for */ 416 int iColumn, /* Column to scan for */ 417 u32 opMask, /* Operator(s) to scan for */ 418 Index *pIdx /* Must be compatible with this index */ 419 ){ 420 pScan->pOrigWC = pWC; 421 pScan->pWC = pWC; 422 pScan->pIdxExpr = 0; 423 pScan->idxaff = 0; 424 pScan->zCollName = 0; 425 pScan->opMask = opMask; 426 pScan->k = 0; 427 pScan->aiCur[0] = iCur; 428 pScan->nEquiv = 1; 429 pScan->iEquiv = 1; 430 if( pIdx ){ 431 int j = iColumn; 432 iColumn = pIdx->aiColumn[j]; 433 if( iColumn==pIdx->pTable->iPKey ){ 434 iColumn = XN_ROWID; 435 }else if( iColumn>=0 ){ 436 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; 437 pScan->zCollName = pIdx->azColl[j]; 438 }else if( iColumn==XN_EXPR ){ 439 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr; 440 pScan->zCollName = pIdx->azColl[j]; 441 pScan->aiColumn[0] = XN_EXPR; 442 return whereScanInitIndexExpr(pScan); 443 } 444 }else if( iColumn==XN_EXPR ){ 445 return 0; 446 } 447 pScan->aiColumn[0] = iColumn; 448 return whereScanNext(pScan); 449 } 450 451 /* 452 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 453 ** where X is a reference to the iColumn of table iCur or of index pIdx 454 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by 455 ** the op parameter. Return a pointer to the term. Return 0 if not found. 456 ** 457 ** If pIdx!=0 then it must be one of the indexes of table iCur. 458 ** Search for terms matching the iColumn-th column of pIdx 459 ** rather than the iColumn-th column of table iCur. 460 ** 461 ** The term returned might by Y=<expr> if there is another constraint in 462 ** the WHERE clause that specifies that X=Y. Any such constraints will be 463 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 464 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11 465 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10 466 ** other equivalent values. Hence a search for X will return <expr> if X=A1 467 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>. 468 ** 469 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 470 ** then try for the one with no dependencies on <expr> - in other words where 471 ** <expr> is a constant expression of some kind. Only return entries of 472 ** the form "X <op> Y" where Y is a column in another table if no terms of 473 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 474 ** exist, try to return a term that does not use WO_EQUIV. 475 */ 476 WhereTerm *sqlite3WhereFindTerm( 477 WhereClause *pWC, /* The WHERE clause to be searched */ 478 int iCur, /* Cursor number of LHS */ 479 int iColumn, /* Column number of LHS */ 480 Bitmask notReady, /* RHS must not overlap with this mask */ 481 u32 op, /* Mask of WO_xx values describing operator */ 482 Index *pIdx /* Must be compatible with this index, if not NULL */ 483 ){ 484 WhereTerm *pResult = 0; 485 WhereTerm *p; 486 WhereScan scan; 487 488 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); 489 op &= WO_EQ|WO_IS; 490 while( p ){ 491 if( (p->prereqRight & notReady)==0 ){ 492 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){ 493 testcase( p->eOperator & WO_IS ); 494 return p; 495 } 496 if( pResult==0 ) pResult = p; 497 } 498 p = whereScanNext(&scan); 499 } 500 return pResult; 501 } 502 503 /* 504 ** This function searches pList for an entry that matches the iCol-th column 505 ** of index pIdx. 506 ** 507 ** If such an expression is found, its index in pList->a[] is returned. If 508 ** no expression is found, -1 is returned. 509 */ 510 static int findIndexCol( 511 Parse *pParse, /* Parse context */ 512 ExprList *pList, /* Expression list to search */ 513 int iBase, /* Cursor for table associated with pIdx */ 514 Index *pIdx, /* Index to match column of */ 515 int iCol /* Column of index to match */ 516 ){ 517 int i; 518 const char *zColl = pIdx->azColl[iCol]; 519 520 for(i=0; i<pList->nExpr; i++){ 521 Expr *p = sqlite3ExprSkipCollateAndLikely(pList->a[i].pExpr); 522 if( ALWAYS(p!=0) 523 && (p->op==TK_COLUMN || p->op==TK_AGG_COLUMN) 524 && p->iColumn==pIdx->aiColumn[iCol] 525 && p->iTable==iBase 526 ){ 527 CollSeq *pColl = sqlite3ExprNNCollSeq(pParse, pList->a[i].pExpr); 528 if( 0==sqlite3StrICmp(pColl->zName, zColl) ){ 529 return i; 530 } 531 } 532 } 533 534 return -1; 535 } 536 537 /* 538 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL 539 */ 540 static int indexColumnNotNull(Index *pIdx, int iCol){ 541 int j; 542 assert( pIdx!=0 ); 543 assert( iCol>=0 && iCol<pIdx->nColumn ); 544 j = pIdx->aiColumn[iCol]; 545 if( j>=0 ){ 546 return pIdx->pTable->aCol[j].notNull; 547 }else if( j==(-1) ){ 548 return 1; 549 }else{ 550 assert( j==(-2) ); 551 return 0; /* Assume an indexed expression can always yield a NULL */ 552 553 } 554 } 555 556 /* 557 ** Return true if the DISTINCT expression-list passed as the third argument 558 ** is redundant. 559 ** 560 ** A DISTINCT list is redundant if any subset of the columns in the 561 ** DISTINCT list are collectively unique and individually non-null. 562 */ 563 static int isDistinctRedundant( 564 Parse *pParse, /* Parsing context */ 565 SrcList *pTabList, /* The FROM clause */ 566 WhereClause *pWC, /* The WHERE clause */ 567 ExprList *pDistinct /* The result set that needs to be DISTINCT */ 568 ){ 569 Table *pTab; 570 Index *pIdx; 571 int i; 572 int iBase; 573 574 /* If there is more than one table or sub-select in the FROM clause of 575 ** this query, then it will not be possible to show that the DISTINCT 576 ** clause is redundant. */ 577 if( pTabList->nSrc!=1 ) return 0; 578 iBase = pTabList->a[0].iCursor; 579 pTab = pTabList->a[0].pTab; 580 581 /* If any of the expressions is an IPK column on table iBase, then return 582 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 583 ** current SELECT is a correlated sub-query. 584 */ 585 for(i=0; i<pDistinct->nExpr; i++){ 586 Expr *p = sqlite3ExprSkipCollateAndLikely(pDistinct->a[i].pExpr); 587 if( NEVER(p==0) ) continue; 588 if( p->op!=TK_COLUMN && p->op!=TK_AGG_COLUMN ) continue; 589 if( p->iTable==iBase && p->iColumn<0 ) return 1; 590 } 591 592 /* Loop through all indices on the table, checking each to see if it makes 593 ** the DISTINCT qualifier redundant. It does so if: 594 ** 595 ** 1. The index is itself UNIQUE, and 596 ** 597 ** 2. All of the columns in the index are either part of the pDistinct 598 ** list, or else the WHERE clause contains a term of the form "col=X", 599 ** where X is a constant value. The collation sequences of the 600 ** comparison and select-list expressions must match those of the index. 601 ** 602 ** 3. All of those index columns for which the WHERE clause does not 603 ** contain a "col=X" term are subject to a NOT NULL constraint. 604 */ 605 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 606 if( !IsUniqueIndex(pIdx) ) continue; 607 if( pIdx->pPartIdxWhere ) continue; 608 for(i=0; i<pIdx->nKeyCol; i++){ 609 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){ 610 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break; 611 if( indexColumnNotNull(pIdx, i)==0 ) break; 612 } 613 } 614 if( i==pIdx->nKeyCol ){ 615 /* This index implies that the DISTINCT qualifier is redundant. */ 616 return 1; 617 } 618 } 619 620 return 0; 621 } 622 623 624 /* 625 ** Estimate the logarithm of the input value to base 2. 626 */ 627 static LogEst estLog(LogEst N){ 628 return N<=10 ? 0 : sqlite3LogEst(N) - 33; 629 } 630 631 /* 632 ** Convert OP_Column opcodes to OP_Copy in previously generated code. 633 ** 634 ** This routine runs over generated VDBE code and translates OP_Column 635 ** opcodes into OP_Copy when the table is being accessed via co-routine 636 ** instead of via table lookup. 637 ** 638 ** If the iAutoidxCur is not zero, then any OP_Rowid instructions on 639 ** cursor iTabCur are transformed into OP_Sequence opcode for the 640 ** iAutoidxCur cursor, in order to generate unique rowids for the 641 ** automatic index being generated. 642 */ 643 static void translateColumnToCopy( 644 Parse *pParse, /* Parsing context */ 645 int iStart, /* Translate from this opcode to the end */ 646 int iTabCur, /* OP_Column/OP_Rowid references to this table */ 647 int iRegister, /* The first column is in this register */ 648 int iAutoidxCur /* If non-zero, cursor of autoindex being generated */ 649 ){ 650 Vdbe *v = pParse->pVdbe; 651 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart); 652 int iEnd = sqlite3VdbeCurrentAddr(v); 653 if( pParse->db->mallocFailed ) return; 654 for(; iStart<iEnd; iStart++, pOp++){ 655 if( pOp->p1!=iTabCur ) continue; 656 if( pOp->opcode==OP_Column ){ 657 pOp->opcode = OP_Copy; 658 pOp->p1 = pOp->p2 + iRegister; 659 pOp->p2 = pOp->p3; 660 pOp->p3 = 0; 661 }else if( pOp->opcode==OP_Rowid ){ 662 pOp->opcode = OP_Sequence; 663 pOp->p1 = iAutoidxCur; 664 #ifdef SQLITE_ALLOW_ROWID_IN_VIEW 665 if( iAutoidxCur==0 ){ 666 pOp->opcode = OP_Null; 667 pOp->p3 = 0; 668 } 669 #endif 670 } 671 } 672 } 673 674 /* 675 ** Two routines for printing the content of an sqlite3_index_info 676 ** structure. Used for testing and debugging only. If neither 677 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 678 ** are no-ops. 679 */ 680 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) 681 static void whereTraceIndexInfoInputs(sqlite3_index_info *p){ 682 int i; 683 if( !sqlite3WhereTrace ) return; 684 for(i=0; i<p->nConstraint; i++){ 685 sqlite3DebugPrintf( 686 " constraint[%d]: col=%d termid=%d op=%d usabled=%d collseq=%s\n", 687 i, 688 p->aConstraint[i].iColumn, 689 p->aConstraint[i].iTermOffset, 690 p->aConstraint[i].op, 691 p->aConstraint[i].usable, 692 sqlite3_vtab_collation(p,i)); 693 } 694 for(i=0; i<p->nOrderBy; i++){ 695 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 696 i, 697 p->aOrderBy[i].iColumn, 698 p->aOrderBy[i].desc); 699 } 700 } 701 static void whereTraceIndexInfoOutputs(sqlite3_index_info *p){ 702 int i; 703 if( !sqlite3WhereTrace ) return; 704 for(i=0; i<p->nConstraint; i++){ 705 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 706 i, 707 p->aConstraintUsage[i].argvIndex, 708 p->aConstraintUsage[i].omit); 709 } 710 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 711 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 712 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 713 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 714 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); 715 } 716 #else 717 #define whereTraceIndexInfoInputs(A) 718 #define whereTraceIndexInfoOutputs(A) 719 #endif 720 721 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 722 /* 723 ** Return TRUE if the WHERE clause term pTerm is of a form where it 724 ** could be used with an index to access pSrc, assuming an appropriate 725 ** index existed. 726 */ 727 static int termCanDriveIndex( 728 const WhereTerm *pTerm, /* WHERE clause term to check */ 729 const SrcItem *pSrc, /* Table we are trying to access */ 730 const Bitmask notReady /* Tables in outer loops of the join */ 731 ){ 732 char aff; 733 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 734 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0; 735 if( (pSrc->fg.jointype & JT_LEFT) 736 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 737 && (pTerm->eOperator & WO_IS) 738 ){ 739 /* Cannot use an IS term from the WHERE clause as an index driver for 740 ** the RHS of a LEFT JOIN. Such a term can only be used if it is from 741 ** the ON clause. */ 742 return 0; 743 } 744 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 745 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 746 if( pTerm->u.x.leftColumn<0 ) return 0; 747 aff = pSrc->pTab->aCol[pTerm->u.x.leftColumn].affinity; 748 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 749 testcase( pTerm->pExpr->op==TK_IS ); 750 return 1; 751 } 752 #endif 753 754 755 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 756 /* 757 ** Generate code to construct the Index object for an automatic index 758 ** and to set up the WhereLevel object pLevel so that the code generator 759 ** makes use of the automatic index. 760 */ 761 static SQLITE_NOINLINE void constructAutomaticIndex( 762 Parse *pParse, /* The parsing context */ 763 const WhereClause *pWC, /* The WHERE clause */ 764 const SrcItem *pSrc, /* The FROM clause term to get the next index */ 765 const Bitmask notReady, /* Mask of cursors that are not available */ 766 WhereLevel *pLevel /* Write new index here */ 767 ){ 768 int nKeyCol; /* Number of columns in the constructed index */ 769 WhereTerm *pTerm; /* A single term of the WHERE clause */ 770 WhereTerm *pWCEnd; /* End of pWC->a[] */ 771 Index *pIdx; /* Object describing the transient index */ 772 Vdbe *v; /* Prepared statement under construction */ 773 int addrInit; /* Address of the initialization bypass jump */ 774 Table *pTable; /* The table being indexed */ 775 int addrTop; /* Top of the index fill loop */ 776 int regRecord; /* Register holding an index record */ 777 int n; /* Column counter */ 778 int i; /* Loop counter */ 779 int mxBitCol; /* Maximum column in pSrc->colUsed */ 780 CollSeq *pColl; /* Collating sequence to on a column */ 781 WhereLoop *pLoop; /* The Loop object */ 782 char *zNotUsed; /* Extra space on the end of pIdx */ 783 Bitmask idxCols; /* Bitmap of columns used for indexing */ 784 Bitmask extraCols; /* Bitmap of additional columns */ 785 u8 sentWarning = 0; /* True if a warnning has been issued */ 786 Expr *pPartial = 0; /* Partial Index Expression */ 787 int iContinue = 0; /* Jump here to skip excluded rows */ 788 SrcItem *pTabItem; /* FROM clause term being indexed */ 789 int addrCounter = 0; /* Address where integer counter is initialized */ 790 int regBase; /* Array of registers where record is assembled */ 791 792 /* Generate code to skip over the creation and initialization of the 793 ** transient index on 2nd and subsequent iterations of the loop. */ 794 v = pParse->pVdbe; 795 assert( v!=0 ); 796 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 797 798 /* Count the number of columns that will be added to the index 799 ** and used to match WHERE clause constraints */ 800 nKeyCol = 0; 801 pTable = pSrc->pTab; 802 pWCEnd = &pWC->a[pWC->nTerm]; 803 pLoop = pLevel->pWLoop; 804 idxCols = 0; 805 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 806 Expr *pExpr = pTerm->pExpr; 807 /* Make the automatic index a partial index if there are terms in the 808 ** WHERE clause (or the ON clause of a LEFT join) that constrain which 809 ** rows of the target table (pSrc) that can be used. */ 810 if( (pTerm->wtFlags & TERM_VIRTUAL)==0 811 && ((pSrc->fg.jointype&JT_LEFT)==0 || ExprHasProperty(pExpr,EP_FromJoin)) 812 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) 813 ){ 814 pPartial = sqlite3ExprAnd(pParse, pPartial, 815 sqlite3ExprDup(pParse->db, pExpr, 0)); 816 } 817 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 818 int iCol; 819 Bitmask cMask; 820 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 821 iCol = pTerm->u.x.leftColumn; 822 cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 823 testcase( iCol==BMS ); 824 testcase( iCol==BMS-1 ); 825 if( !sentWarning ){ 826 sqlite3_log(SQLITE_WARNING_AUTOINDEX, 827 "automatic index on %s(%s)", pTable->zName, 828 pTable->aCol[iCol].zCnName); 829 sentWarning = 1; 830 } 831 if( (idxCols & cMask)==0 ){ 832 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ 833 goto end_auto_index_create; 834 } 835 pLoop->aLTerm[nKeyCol++] = pTerm; 836 idxCols |= cMask; 837 } 838 } 839 } 840 assert( nKeyCol>0 || pParse->db->mallocFailed ); 841 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; 842 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED 843 | WHERE_AUTO_INDEX; 844 845 /* Count the number of additional columns needed to create a 846 ** covering index. A "covering index" is an index that contains all 847 ** columns that are needed by the query. With a covering index, the 848 ** original table never needs to be accessed. Automatic indices must 849 ** be a covering index because the index will not be updated if the 850 ** original table changes and the index and table cannot both be used 851 ** if they go out of sync. 852 */ 853 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); 854 mxBitCol = MIN(BMS-1,pTable->nCol); 855 testcase( pTable->nCol==BMS-1 ); 856 testcase( pTable->nCol==BMS-2 ); 857 for(i=0; i<mxBitCol; i++){ 858 if( extraCols & MASKBIT(i) ) nKeyCol++; 859 } 860 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 861 nKeyCol += pTable->nCol - BMS + 1; 862 } 863 864 /* Construct the Index object to describe this index */ 865 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); 866 if( pIdx==0 ) goto end_auto_index_create; 867 pLoop->u.btree.pIndex = pIdx; 868 pIdx->zName = "auto-index"; 869 pIdx->pTable = pTable; 870 n = 0; 871 idxCols = 0; 872 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 873 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 874 int iCol; 875 Bitmask cMask; 876 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 877 iCol = pTerm->u.x.leftColumn; 878 cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 879 testcase( iCol==BMS-1 ); 880 testcase( iCol==BMS ); 881 if( (idxCols & cMask)==0 ){ 882 Expr *pX = pTerm->pExpr; 883 idxCols |= cMask; 884 pIdx->aiColumn[n] = pTerm->u.x.leftColumn; 885 pColl = sqlite3ExprCompareCollSeq(pParse, pX); 886 assert( pColl!=0 || pParse->nErr>0 ); /* TH3 collate01.800 */ 887 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY; 888 n++; 889 } 890 } 891 } 892 assert( (u32)n==pLoop->u.btree.nEq ); 893 894 /* Add additional columns needed to make the automatic index into 895 ** a covering index */ 896 for(i=0; i<mxBitCol; i++){ 897 if( extraCols & MASKBIT(i) ){ 898 pIdx->aiColumn[n] = i; 899 pIdx->azColl[n] = sqlite3StrBINARY; 900 n++; 901 } 902 } 903 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 904 for(i=BMS-1; i<pTable->nCol; i++){ 905 pIdx->aiColumn[n] = i; 906 pIdx->azColl[n] = sqlite3StrBINARY; 907 n++; 908 } 909 } 910 assert( n==nKeyCol ); 911 pIdx->aiColumn[n] = XN_ROWID; 912 pIdx->azColl[n] = sqlite3StrBINARY; 913 914 /* Create the automatic index */ 915 assert( pLevel->iIdxCur>=0 ); 916 pLevel->iIdxCur = pParse->nTab++; 917 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); 918 sqlite3VdbeSetP4KeyInfo(pParse, pIdx); 919 VdbeComment((v, "for %s", pTable->zName)); 920 if( OptimizationEnabled(pParse->db, SQLITE_BloomFilter) ){ 921 pLevel->regFilter = ++pParse->nMem; 922 sqlite3VdbeAddOp2(v, OP_Blob, 10000, pLevel->regFilter); 923 } 924 925 /* Fill the automatic index with content */ 926 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom]; 927 if( pTabItem->fg.viaCoroutine ){ 928 int regYield = pTabItem->regReturn; 929 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0); 930 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); 931 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 932 VdbeCoverage(v); 933 VdbeComment((v, "next row of %s", pTabItem->pTab->zName)); 934 }else{ 935 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); 936 } 937 if( pPartial ){ 938 iContinue = sqlite3VdbeMakeLabel(pParse); 939 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); 940 pLoop->wsFlags |= WHERE_PARTIALIDX; 941 } 942 regRecord = sqlite3GetTempReg(pParse); 943 regBase = sqlite3GenerateIndexKey( 944 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0 945 ); 946 if( pLevel->regFilter ){ 947 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0, 948 regBase, pLoop->u.btree.nEq); 949 } 950 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 951 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 952 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); 953 if( pTabItem->fg.viaCoroutine ){ 954 sqlite3VdbeChangeP2(v, addrCounter, regBase+n); 955 testcase( pParse->db->mallocFailed ); 956 assert( pLevel->iIdxCur>0 ); 957 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur, 958 pTabItem->regResult, pLevel->iIdxCur); 959 sqlite3VdbeGoto(v, addrTop); 960 pTabItem->fg.viaCoroutine = 0; 961 }else{ 962 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); 963 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 964 } 965 sqlite3VdbeJumpHere(v, addrTop); 966 sqlite3ReleaseTempReg(pParse, regRecord); 967 968 /* Jump here when skipping the initialization */ 969 sqlite3VdbeJumpHere(v, addrInit); 970 971 end_auto_index_create: 972 sqlite3ExprDelete(pParse->db, pPartial); 973 } 974 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 975 976 /* 977 ** Generate bytecode that will initialize a Bloom filter that is appropriate 978 ** for pLevel. 979 ** 980 ** If there are inner loops within pLevel that have the WHERE_BLOOMFILTER 981 ** flag set, initialize a Bloomfilter for them as well. Except don't do 982 ** this recursive initialization if the SQLITE_BloomPulldown optimization has 983 ** been turned off. 984 ** 985 ** When the Bloom filter is initialized, the WHERE_BLOOMFILTER flag is cleared 986 ** from the loop, but the regFilter value is set to a register that implements 987 ** the Bloom filter. When regFilter is positive, the 988 ** sqlite3WhereCodeOneLoopStart() will generate code to test the Bloom filter 989 ** and skip the subsequence B-Tree seek if the Bloom filter indicates that 990 ** no matching rows exist. 991 ** 992 ** This routine may only be called if it has previously been determined that 993 ** the loop would benefit from a Bloom filter, and the WHERE_BLOOMFILTER bit 994 ** is set. 995 */ 996 static SQLITE_NOINLINE void sqlite3ConstructBloomFilter( 997 WhereInfo *pWInfo, /* The WHERE clause */ 998 int iLevel, /* Index in pWInfo->a[] that is pLevel */ 999 WhereLevel *pLevel, /* Make a Bloom filter for this FROM term */ 1000 Bitmask notReady /* Loops that are not ready */ 1001 ){ 1002 int addrOnce; /* Address of opening OP_Once */ 1003 int addrTop; /* Address of OP_Rewind */ 1004 int addrCont; /* Jump here to skip a row */ 1005 const WhereTerm *pTerm; /* For looping over WHERE clause terms */ 1006 const WhereTerm *pWCEnd; /* Last WHERE clause term */ 1007 Parse *pParse = pWInfo->pParse; /* Parsing context */ 1008 Vdbe *v = pParse->pVdbe; /* VDBE under construction */ 1009 WhereLoop *pLoop = pLevel->pWLoop; /* The loop being coded */ 1010 int iCur; /* Cursor for table getting the filter */ 1011 1012 assert( pLoop!=0 ); 1013 assert( v!=0 ); 1014 assert( pLoop->wsFlags & WHERE_BLOOMFILTER ); 1015 1016 addrOnce = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 1017 do{ 1018 const SrcItem *pItem; 1019 const Table *pTab; 1020 u64 sz; 1021 sqlite3WhereExplainBloomFilter(pParse, pWInfo, pLevel); 1022 addrCont = sqlite3VdbeMakeLabel(pParse); 1023 iCur = pLevel->iTabCur; 1024 pLevel->regFilter = ++pParse->nMem; 1025 1026 /* The Bloom filter is a Blob held in a register. Initialize it 1027 ** to zero-filled blob of at least 80K bits, but maybe more if the 1028 ** estimated size of the table is larger. We could actually 1029 ** measure the size of the table at run-time using OP_Count with 1030 ** P3==1 and use that value to initialize the blob. But that makes 1031 ** testing complicated. By basing the blob size on the value in the 1032 ** sqlite_stat1 table, testing is much easier. 1033 */ 1034 pItem = &pWInfo->pTabList->a[pLevel->iFrom]; 1035 assert( pItem!=0 ); 1036 pTab = pItem->pTab; 1037 assert( pTab!=0 ); 1038 sz = sqlite3LogEstToInt(pTab->nRowLogEst); 1039 if( sz<10000 ){ 1040 sz = 10000; 1041 }else if( sz>10000000 ){ 1042 sz = 10000000; 1043 } 1044 sqlite3VdbeAddOp2(v, OP_Blob, (int)sz, pLevel->regFilter); 1045 1046 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, iCur); VdbeCoverage(v); 1047 pWCEnd = &pWInfo->sWC.a[pWInfo->sWC.nTerm]; 1048 for(pTerm=pWInfo->sWC.a; pTerm<pWCEnd; pTerm++){ 1049 Expr *pExpr = pTerm->pExpr; 1050 if( (pTerm->wtFlags & TERM_VIRTUAL)==0 1051 && sqlite3ExprIsTableConstant(pExpr, iCur) 1052 ){ 1053 sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL); 1054 } 1055 } 1056 if( pLoop->wsFlags & WHERE_IPK ){ 1057 int r1 = sqlite3GetTempReg(pParse); 1058 sqlite3VdbeAddOp2(v, OP_Rowid, iCur, r1); 1059 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0, r1, 1); 1060 sqlite3ReleaseTempReg(pParse, r1); 1061 }else{ 1062 Index *pIdx = pLoop->u.btree.pIndex; 1063 int n = pLoop->u.btree.nEq; 1064 int r1 = sqlite3GetTempRange(pParse, n); 1065 int jj; 1066 for(jj=0; jj<n; jj++){ 1067 int iCol = pIdx->aiColumn[jj]; 1068 assert( pIdx->pTable==pItem->pTab ); 1069 sqlite3ExprCodeGetColumnOfTable(v, pIdx->pTable, iCur, iCol,r1+jj); 1070 } 1071 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0, r1, n); 1072 sqlite3ReleaseTempRange(pParse, r1, n); 1073 } 1074 sqlite3VdbeResolveLabel(v, addrCont); 1075 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); 1076 VdbeCoverage(v); 1077 sqlite3VdbeJumpHere(v, addrTop); 1078 pLoop->wsFlags &= ~WHERE_BLOOMFILTER; 1079 if( OptimizationDisabled(pParse->db, SQLITE_BloomPulldown) ) break; 1080 while( ++iLevel < pWInfo->nLevel ){ 1081 const SrcItem *pTabItem; 1082 pLevel = &pWInfo->a[iLevel]; 1083 pTabItem = &pWInfo->pTabList->a[pLevel->iFrom]; 1084 if( pTabItem->fg.jointype & JT_LEFT ) continue; 1085 pLoop = pLevel->pWLoop; 1086 if( NEVER(pLoop==0) ) continue; 1087 if( pLoop->prereq & notReady ) continue; 1088 if( (pLoop->wsFlags & (WHERE_BLOOMFILTER|WHERE_COLUMN_IN)) 1089 ==WHERE_BLOOMFILTER 1090 ){ 1091 /* This is a candidate for bloom-filter pull-down (early evaluation). 1092 ** The test that WHERE_COLUMN_IN is omitted is important, as we are 1093 ** not able to do early evaluation of bloom filters that make use of 1094 ** the IN operator */ 1095 break; 1096 } 1097 } 1098 }while( iLevel < pWInfo->nLevel ); 1099 sqlite3VdbeJumpHere(v, addrOnce); 1100 } 1101 1102 1103 #ifndef SQLITE_OMIT_VIRTUALTABLE 1104 /* 1105 ** Allocate and populate an sqlite3_index_info structure. It is the 1106 ** responsibility of the caller to eventually release the structure 1107 ** by passing the pointer returned by this function to freeIndexInfo(). 1108 */ 1109 static sqlite3_index_info *allocateIndexInfo( 1110 WhereInfo *pWInfo, /* The WHERE clause */ 1111 WhereClause *pWC, /* The WHERE clause being analyzed */ 1112 Bitmask mUnusable, /* Ignore terms with these prereqs */ 1113 SrcItem *pSrc, /* The FROM clause term that is the vtab */ 1114 u16 *pmNoOmit /* Mask of terms not to omit */ 1115 ){ 1116 int i, j; 1117 int nTerm; 1118 Parse *pParse = pWInfo->pParse; 1119 struct sqlite3_index_constraint *pIdxCons; 1120 struct sqlite3_index_orderby *pIdxOrderBy; 1121 struct sqlite3_index_constraint_usage *pUsage; 1122 struct HiddenIndexInfo *pHidden; 1123 WhereTerm *pTerm; 1124 int nOrderBy; 1125 sqlite3_index_info *pIdxInfo; 1126 u16 mNoOmit = 0; 1127 const Table *pTab; 1128 int eDistinct = 0; 1129 ExprList *pOrderBy = pWInfo->pOrderBy; 1130 1131 assert( pSrc!=0 ); 1132 pTab = pSrc->pTab; 1133 assert( pTab!=0 ); 1134 assert( IsVirtual(pTab) ); 1135 1136 /* Find all WHERE clause constraints referring to this virtual table. 1137 ** Mark each term with the TERM_OK flag. Set nTerm to the number of 1138 ** terms found. 1139 */ 1140 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 1141 pTerm->wtFlags &= ~TERM_OK; 1142 if( pTerm->leftCursor != pSrc->iCursor ) continue; 1143 if( pTerm->prereqRight & mUnusable ) continue; 1144 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 1145 testcase( pTerm->eOperator & WO_IN ); 1146 testcase( pTerm->eOperator & WO_ISNULL ); 1147 testcase( pTerm->eOperator & WO_IS ); 1148 testcase( pTerm->eOperator & WO_ALL ); 1149 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 1150 if( pTerm->wtFlags & TERM_VNULL ) continue; 1151 1152 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 1153 assert( pTerm->u.x.leftColumn>=XN_ROWID ); 1154 assert( pTerm->u.x.leftColumn<pTab->nCol ); 1155 1156 /* tag-20191211-002: WHERE-clause constraints are not useful to the 1157 ** right-hand table of a LEFT JOIN. See tag-20191211-001 for the 1158 ** equivalent restriction for ordinary tables. */ 1159 if( (pSrc->fg.jointype & JT_LEFT)!=0 1160 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 1161 ){ 1162 continue; 1163 } 1164 nTerm++; 1165 pTerm->wtFlags |= TERM_OK; 1166 } 1167 1168 /* If the ORDER BY clause contains only columns in the current 1169 ** virtual table then allocate space for the aOrderBy part of 1170 ** the sqlite3_index_info structure. 1171 */ 1172 nOrderBy = 0; 1173 if( pOrderBy ){ 1174 int n = pOrderBy->nExpr; 1175 for(i=0; i<n; i++){ 1176 Expr *pExpr = pOrderBy->a[i].pExpr; 1177 Expr *pE2; 1178 1179 /* Skip over constant terms in the ORDER BY clause */ 1180 if( sqlite3ExprIsConstant(pExpr) ){ 1181 continue; 1182 } 1183 1184 /* Virtual tables are unable to deal with NULLS FIRST */ 1185 if( pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL ) break; 1186 1187 /* First case - a direct column references without a COLLATE operator */ 1188 if( pExpr->op==TK_COLUMN && pExpr->iTable==pSrc->iCursor ){ 1189 assert( pExpr->iColumn>=XN_ROWID && pExpr->iColumn<pTab->nCol ); 1190 continue; 1191 } 1192 1193 /* 2nd case - a column reference with a COLLATE operator. Only match 1194 ** of the COLLATE operator matches the collation of the column. */ 1195 if( pExpr->op==TK_COLLATE 1196 && (pE2 = pExpr->pLeft)->op==TK_COLUMN 1197 && pE2->iTable==pSrc->iCursor 1198 ){ 1199 const char *zColl; /* The collating sequence name */ 1200 assert( !ExprHasProperty(pExpr, EP_IntValue) ); 1201 assert( pExpr->u.zToken!=0 ); 1202 assert( pE2->iColumn>=XN_ROWID && pE2->iColumn<pTab->nCol ); 1203 pExpr->iColumn = pE2->iColumn; 1204 if( pE2->iColumn<0 ) continue; /* Collseq does not matter for rowid */ 1205 zColl = sqlite3ColumnColl(&pTab->aCol[pE2->iColumn]); 1206 if( zColl==0 ) zColl = sqlite3StrBINARY; 1207 if( sqlite3_stricmp(pExpr->u.zToken, zColl)==0 ) continue; 1208 } 1209 1210 /* No matches cause a break out of the loop */ 1211 break; 1212 } 1213 if( i==n ){ 1214 nOrderBy = n; 1215 if( (pWInfo->wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY)) ){ 1216 eDistinct = 1 + ((pWInfo->wctrlFlags & WHERE_DISTINCTBY)!=0); 1217 } 1218 } 1219 } 1220 1221 /* Allocate the sqlite3_index_info structure 1222 */ 1223 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 1224 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 1225 + sizeof(*pIdxOrderBy)*nOrderBy + sizeof(*pHidden) 1226 + sizeof(sqlite3_value*)*nTerm ); 1227 if( pIdxInfo==0 ){ 1228 sqlite3ErrorMsg(pParse, "out of memory"); 1229 return 0; 1230 } 1231 pHidden = (struct HiddenIndexInfo*)&pIdxInfo[1]; 1232 pIdxCons = (struct sqlite3_index_constraint*)&pHidden->aRhs[nTerm]; 1233 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 1234 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 1235 pIdxInfo->aConstraint = pIdxCons; 1236 pIdxInfo->aOrderBy = pIdxOrderBy; 1237 pIdxInfo->aConstraintUsage = pUsage; 1238 pHidden->pWC = pWC; 1239 pHidden->pParse = pParse; 1240 pHidden->eDistinct = eDistinct; 1241 pHidden->mIn = 0; 1242 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 1243 u16 op; 1244 if( (pTerm->wtFlags & TERM_OK)==0 ) continue; 1245 pIdxCons[j].iColumn = pTerm->u.x.leftColumn; 1246 pIdxCons[j].iTermOffset = i; 1247 op = pTerm->eOperator & WO_ALL; 1248 if( op==WO_IN ){ 1249 if( (pTerm->wtFlags & TERM_SLICE)==0 ){ 1250 pHidden->mIn |= SMASKBIT32(j); 1251 } 1252 op = WO_EQ; 1253 } 1254 if( op==WO_AUX ){ 1255 pIdxCons[j].op = pTerm->eMatchOp; 1256 }else if( op & (WO_ISNULL|WO_IS) ){ 1257 if( op==WO_ISNULL ){ 1258 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_ISNULL; 1259 }else{ 1260 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_IS; 1261 } 1262 }else{ 1263 pIdxCons[j].op = (u8)op; 1264 /* The direct assignment in the previous line is possible only because 1265 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 1266 ** following asserts verify this fact. */ 1267 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 1268 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 1269 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 1270 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 1271 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 1272 assert( pTerm->eOperator&(WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_AUX) ); 1273 1274 if( op & (WO_LT|WO_LE|WO_GT|WO_GE) 1275 && sqlite3ExprIsVector(pTerm->pExpr->pRight) 1276 ){ 1277 testcase( j!=i ); 1278 if( j<16 ) mNoOmit |= (1 << j); 1279 if( op==WO_LT ) pIdxCons[j].op = WO_LE; 1280 if( op==WO_GT ) pIdxCons[j].op = WO_GE; 1281 } 1282 } 1283 1284 j++; 1285 } 1286 assert( j==nTerm ); 1287 pIdxInfo->nConstraint = j; 1288 for(i=j=0; i<nOrderBy; i++){ 1289 Expr *pExpr = pOrderBy->a[i].pExpr; 1290 if( sqlite3ExprIsConstant(pExpr) ) continue; 1291 assert( pExpr->op==TK_COLUMN 1292 || (pExpr->op==TK_COLLATE && pExpr->pLeft->op==TK_COLUMN 1293 && pExpr->iColumn==pExpr->pLeft->iColumn) ); 1294 pIdxOrderBy[j].iColumn = pExpr->iColumn; 1295 pIdxOrderBy[j].desc = pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC; 1296 j++; 1297 } 1298 pIdxInfo->nOrderBy = j; 1299 1300 *pmNoOmit = mNoOmit; 1301 return pIdxInfo; 1302 } 1303 1304 /* 1305 ** Free an sqlite3_index_info structure allocated by allocateIndexInfo() 1306 ** and possibly modified by xBestIndex methods. 1307 */ 1308 static void freeIndexInfo(sqlite3 *db, sqlite3_index_info *pIdxInfo){ 1309 HiddenIndexInfo *pHidden; 1310 int i; 1311 assert( pIdxInfo!=0 ); 1312 pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 1313 assert( pHidden->pParse!=0 ); 1314 assert( pHidden->pParse->db==db ); 1315 for(i=0; i<pIdxInfo->nConstraint; i++){ 1316 sqlite3ValueFree(pHidden->aRhs[i]); /* IMP: R-14553-25174 */ 1317 pHidden->aRhs[i] = 0; 1318 } 1319 sqlite3DbFree(db, pIdxInfo); 1320 } 1321 1322 /* 1323 ** The table object reference passed as the second argument to this function 1324 ** must represent a virtual table. This function invokes the xBestIndex() 1325 ** method of the virtual table with the sqlite3_index_info object that 1326 ** comes in as the 3rd argument to this function. 1327 ** 1328 ** If an error occurs, pParse is populated with an error message and an 1329 ** appropriate error code is returned. A return of SQLITE_CONSTRAINT from 1330 ** xBestIndex is not considered an error. SQLITE_CONSTRAINT indicates that 1331 ** the current configuration of "unusable" flags in sqlite3_index_info can 1332 ** not result in a valid plan. 1333 ** 1334 ** Whether or not an error is returned, it is the responsibility of the 1335 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 1336 ** that this is required. 1337 */ 1338 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 1339 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 1340 int rc; 1341 1342 whereTraceIndexInfoInputs(p); 1343 pParse->db->nSchemaLock++; 1344 rc = pVtab->pModule->xBestIndex(pVtab, p); 1345 pParse->db->nSchemaLock--; 1346 whereTraceIndexInfoOutputs(p); 1347 1348 if( rc!=SQLITE_OK && rc!=SQLITE_CONSTRAINT ){ 1349 if( rc==SQLITE_NOMEM ){ 1350 sqlite3OomFault(pParse->db); 1351 }else if( !pVtab->zErrMsg ){ 1352 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 1353 }else{ 1354 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 1355 } 1356 } 1357 sqlite3_free(pVtab->zErrMsg); 1358 pVtab->zErrMsg = 0; 1359 return rc; 1360 } 1361 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 1362 1363 #ifdef SQLITE_ENABLE_STAT4 1364 /* 1365 ** Estimate the location of a particular key among all keys in an 1366 ** index. Store the results in aStat as follows: 1367 ** 1368 ** aStat[0] Est. number of rows less than pRec 1369 ** aStat[1] Est. number of rows equal to pRec 1370 ** 1371 ** Return the index of the sample that is the smallest sample that 1372 ** is greater than or equal to pRec. Note that this index is not an index 1373 ** into the aSample[] array - it is an index into a virtual set of samples 1374 ** based on the contents of aSample[] and the number of fields in record 1375 ** pRec. 1376 */ 1377 static int whereKeyStats( 1378 Parse *pParse, /* Database connection */ 1379 Index *pIdx, /* Index to consider domain of */ 1380 UnpackedRecord *pRec, /* Vector of values to consider */ 1381 int roundUp, /* Round up if true. Round down if false */ 1382 tRowcnt *aStat /* OUT: stats written here */ 1383 ){ 1384 IndexSample *aSample = pIdx->aSample; 1385 int iCol; /* Index of required stats in anEq[] etc. */ 1386 int i; /* Index of first sample >= pRec */ 1387 int iSample; /* Smallest sample larger than or equal to pRec */ 1388 int iMin = 0; /* Smallest sample not yet tested */ 1389 int iTest; /* Next sample to test */ 1390 int res; /* Result of comparison operation */ 1391 int nField; /* Number of fields in pRec */ 1392 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 1393 1394 #ifndef SQLITE_DEBUG 1395 UNUSED_PARAMETER( pParse ); 1396 #endif 1397 assert( pRec!=0 ); 1398 assert( pIdx->nSample>0 ); 1399 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1400 1401 /* Do a binary search to find the first sample greater than or equal 1402 ** to pRec. If pRec contains a single field, the set of samples to search 1403 ** is simply the aSample[] array. If the samples in aSample[] contain more 1404 ** than one fields, all fields following the first are ignored. 1405 ** 1406 ** If pRec contains N fields, where N is more than one, then as well as the 1407 ** samples in aSample[] (truncated to N fields), the search also has to 1408 ** consider prefixes of those samples. For example, if the set of samples 1409 ** in aSample is: 1410 ** 1411 ** aSample[0] = (a, 5) 1412 ** aSample[1] = (a, 10) 1413 ** aSample[2] = (b, 5) 1414 ** aSample[3] = (c, 100) 1415 ** aSample[4] = (c, 105) 1416 ** 1417 ** Then the search space should ideally be the samples above and the 1418 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1419 ** the code actually searches this set: 1420 ** 1421 ** 0: (a) 1422 ** 1: (a, 5) 1423 ** 2: (a, 10) 1424 ** 3: (a, 10) 1425 ** 4: (b) 1426 ** 5: (b, 5) 1427 ** 6: (c) 1428 ** 7: (c, 100) 1429 ** 8: (c, 105) 1430 ** 9: (c, 105) 1431 ** 1432 ** For each sample in the aSample[] array, N samples are present in the 1433 ** effective sample array. In the above, samples 0 and 1 are based on 1434 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1435 ** 1436 ** Often, sample i of each block of N effective samples has (i+1) fields. 1437 ** Except, each sample may be extended to ensure that it is greater than or 1438 ** equal to the previous sample in the array. For example, in the above, 1439 ** sample 2 is the first sample of a block of N samples, so at first it 1440 ** appears that it should be 1 field in size. However, that would make it 1441 ** smaller than sample 1, so the binary search would not work. As a result, 1442 ** it is extended to two fields. The duplicates that this creates do not 1443 ** cause any problems. 1444 */ 1445 nField = pRec->nField; 1446 iCol = 0; 1447 iSample = pIdx->nSample * nField; 1448 do{ 1449 int iSamp; /* Index in aSample[] of test sample */ 1450 int n; /* Number of fields in test sample */ 1451 1452 iTest = (iMin+iSample)/2; 1453 iSamp = iTest / nField; 1454 if( iSamp>0 ){ 1455 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1456 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1457 ** fields that is greater than the previous effective sample. */ 1458 for(n=(iTest % nField) + 1; n<nField; n++){ 1459 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1460 } 1461 }else{ 1462 n = iTest + 1; 1463 } 1464 1465 pRec->nField = n; 1466 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1467 if( res<0 ){ 1468 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1469 iMin = iTest+1; 1470 }else if( res==0 && n<nField ){ 1471 iLower = aSample[iSamp].anLt[n-1]; 1472 iMin = iTest+1; 1473 res = -1; 1474 }else{ 1475 iSample = iTest; 1476 iCol = n-1; 1477 } 1478 }while( res && iMin<iSample ); 1479 i = iSample / nField; 1480 1481 #ifdef SQLITE_DEBUG 1482 /* The following assert statements check that the binary search code 1483 ** above found the right answer. This block serves no purpose other 1484 ** than to invoke the asserts. */ 1485 if( pParse->db->mallocFailed==0 ){ 1486 if( res==0 ){ 1487 /* If (res==0) is true, then pRec must be equal to sample i. */ 1488 assert( i<pIdx->nSample ); 1489 assert( iCol==nField-1 ); 1490 pRec->nField = nField; 1491 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1492 || pParse->db->mallocFailed 1493 ); 1494 }else{ 1495 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1496 ** all samples in the aSample[] array, pRec must be smaller than the 1497 ** (iCol+1) field prefix of sample i. */ 1498 assert( i<=pIdx->nSample && i>=0 ); 1499 pRec->nField = iCol+1; 1500 assert( i==pIdx->nSample 1501 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1502 || pParse->db->mallocFailed ); 1503 1504 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1505 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1506 ** be greater than or equal to the (iCol) field prefix of sample i. 1507 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1508 if( iCol>0 ){ 1509 pRec->nField = iCol; 1510 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1511 || pParse->db->mallocFailed ); 1512 } 1513 if( i>0 ){ 1514 pRec->nField = nField; 1515 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1516 || pParse->db->mallocFailed ); 1517 } 1518 } 1519 } 1520 #endif /* ifdef SQLITE_DEBUG */ 1521 1522 if( res==0 ){ 1523 /* Record pRec is equal to sample i */ 1524 assert( iCol==nField-1 ); 1525 aStat[0] = aSample[i].anLt[iCol]; 1526 aStat[1] = aSample[i].anEq[iCol]; 1527 }else{ 1528 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1529 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1530 ** is larger than all samples in the array. */ 1531 tRowcnt iUpper, iGap; 1532 if( i>=pIdx->nSample ){ 1533 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1534 }else{ 1535 iUpper = aSample[i].anLt[iCol]; 1536 } 1537 1538 if( iLower>=iUpper ){ 1539 iGap = 0; 1540 }else{ 1541 iGap = iUpper - iLower; 1542 } 1543 if( roundUp ){ 1544 iGap = (iGap*2)/3; 1545 }else{ 1546 iGap = iGap/3; 1547 } 1548 aStat[0] = iLower + iGap; 1549 aStat[1] = pIdx->aAvgEq[nField-1]; 1550 } 1551 1552 /* Restore the pRec->nField value before returning. */ 1553 pRec->nField = nField; 1554 return i; 1555 } 1556 #endif /* SQLITE_ENABLE_STAT4 */ 1557 1558 /* 1559 ** If it is not NULL, pTerm is a term that provides an upper or lower 1560 ** bound on a range scan. Without considering pTerm, it is estimated 1561 ** that the scan will visit nNew rows. This function returns the number 1562 ** estimated to be visited after taking pTerm into account. 1563 ** 1564 ** If the user explicitly specified a likelihood() value for this term, 1565 ** then the return value is the likelihood multiplied by the number of 1566 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1567 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1568 */ 1569 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1570 LogEst nRet = nNew; 1571 if( pTerm ){ 1572 if( pTerm->truthProb<=0 ){ 1573 nRet += pTerm->truthProb; 1574 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1575 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1576 } 1577 } 1578 return nRet; 1579 } 1580 1581 1582 #ifdef SQLITE_ENABLE_STAT4 1583 /* 1584 ** Return the affinity for a single column of an index. 1585 */ 1586 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1587 assert( iCol>=0 && iCol<pIdx->nColumn ); 1588 if( !pIdx->zColAff ){ 1589 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1590 } 1591 assert( pIdx->zColAff[iCol]!=0 ); 1592 return pIdx->zColAff[iCol]; 1593 } 1594 #endif 1595 1596 1597 #ifdef SQLITE_ENABLE_STAT4 1598 /* 1599 ** This function is called to estimate the number of rows visited by a 1600 ** range-scan on a skip-scan index. For example: 1601 ** 1602 ** CREATE INDEX i1 ON t1(a, b, c); 1603 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1604 ** 1605 ** Value pLoop->nOut is currently set to the estimated number of rows 1606 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1607 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1608 ** on the stat4 data for the index. this scan will be peformed multiple 1609 ** times (once for each (a,b) combination that matches a=?) is dealt with 1610 ** by the caller. 1611 ** 1612 ** It does this by scanning through all stat4 samples, comparing values 1613 ** extracted from pLower and pUpper with the corresponding column in each 1614 ** sample. If L and U are the number of samples found to be less than or 1615 ** equal to the values extracted from pLower and pUpper respectively, and 1616 ** N is the total number of samples, the pLoop->nOut value is adjusted 1617 ** as follows: 1618 ** 1619 ** nOut = nOut * ( min(U - L, 1) / N ) 1620 ** 1621 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1622 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1623 ** U is set to N. 1624 ** 1625 ** Normally, this function sets *pbDone to 1 before returning. However, 1626 ** if no value can be extracted from either pLower or pUpper (and so the 1627 ** estimate of the number of rows delivered remains unchanged), *pbDone 1628 ** is left as is. 1629 ** 1630 ** If an error occurs, an SQLite error code is returned. Otherwise, 1631 ** SQLITE_OK. 1632 */ 1633 static int whereRangeSkipScanEst( 1634 Parse *pParse, /* Parsing & code generating context */ 1635 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1636 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1637 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1638 int *pbDone /* Set to true if at least one expr. value extracted */ 1639 ){ 1640 Index *p = pLoop->u.btree.pIndex; 1641 int nEq = pLoop->u.btree.nEq; 1642 sqlite3 *db = pParse->db; 1643 int nLower = -1; 1644 int nUpper = p->nSample+1; 1645 int rc = SQLITE_OK; 1646 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1647 CollSeq *pColl; 1648 1649 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1650 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1651 sqlite3_value *pVal = 0; /* Value extracted from record */ 1652 1653 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1654 if( pLower ){ 1655 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1656 nLower = 0; 1657 } 1658 if( pUpper && rc==SQLITE_OK ){ 1659 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1660 nUpper = p2 ? 0 : p->nSample; 1661 } 1662 1663 if( p1 || p2 ){ 1664 int i; 1665 int nDiff; 1666 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1667 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1668 if( rc==SQLITE_OK && p1 ){ 1669 int res = sqlite3MemCompare(p1, pVal, pColl); 1670 if( res>=0 ) nLower++; 1671 } 1672 if( rc==SQLITE_OK && p2 ){ 1673 int res = sqlite3MemCompare(p2, pVal, pColl); 1674 if( res>=0 ) nUpper++; 1675 } 1676 } 1677 nDiff = (nUpper - nLower); 1678 if( nDiff<=0 ) nDiff = 1; 1679 1680 /* If there is both an upper and lower bound specified, and the 1681 ** comparisons indicate that they are close together, use the fallback 1682 ** method (assume that the scan visits 1/64 of the rows) for estimating 1683 ** the number of rows visited. Otherwise, estimate the number of rows 1684 ** using the method described in the header comment for this function. */ 1685 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1686 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1687 pLoop->nOut -= nAdjust; 1688 *pbDone = 1; 1689 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1690 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1691 } 1692 1693 }else{ 1694 assert( *pbDone==0 ); 1695 } 1696 1697 sqlite3ValueFree(p1); 1698 sqlite3ValueFree(p2); 1699 sqlite3ValueFree(pVal); 1700 1701 return rc; 1702 } 1703 #endif /* SQLITE_ENABLE_STAT4 */ 1704 1705 /* 1706 ** This function is used to estimate the number of rows that will be visited 1707 ** by scanning an index for a range of values. The range may have an upper 1708 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1709 ** and lower bounds are represented by pLower and pUpper respectively. For 1710 ** example, assuming that index p is on t1(a): 1711 ** 1712 ** ... FROM t1 WHERE a > ? AND a < ? ... 1713 ** |_____| |_____| 1714 ** | | 1715 ** pLower pUpper 1716 ** 1717 ** If either of the upper or lower bound is not present, then NULL is passed in 1718 ** place of the corresponding WhereTerm. 1719 ** 1720 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1721 ** column subject to the range constraint. Or, equivalently, the number of 1722 ** equality constraints optimized by the proposed index scan. For example, 1723 ** assuming index p is on t1(a, b), and the SQL query is: 1724 ** 1725 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1726 ** 1727 ** then nEq is set to 1 (as the range restricted column, b, is the second 1728 ** left-most column of the index). Or, if the query is: 1729 ** 1730 ** ... FROM t1 WHERE a > ? AND a < ? ... 1731 ** 1732 ** then nEq is set to 0. 1733 ** 1734 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1735 ** number of rows that the index scan is expected to visit without 1736 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1737 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1738 ** to account for the range constraints pLower and pUpper. 1739 ** 1740 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1741 ** used, a single range inequality reduces the search space by a factor of 4. 1742 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1743 ** rows visited by a factor of 64. 1744 */ 1745 static int whereRangeScanEst( 1746 Parse *pParse, /* Parsing & code generating context */ 1747 WhereLoopBuilder *pBuilder, 1748 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1749 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1750 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1751 ){ 1752 int rc = SQLITE_OK; 1753 int nOut = pLoop->nOut; 1754 LogEst nNew; 1755 1756 #ifdef SQLITE_ENABLE_STAT4 1757 Index *p = pLoop->u.btree.pIndex; 1758 int nEq = pLoop->u.btree.nEq; 1759 1760 if( p->nSample>0 && ALWAYS(nEq<p->nSampleCol) 1761 && OptimizationEnabled(pParse->db, SQLITE_Stat4) 1762 ){ 1763 if( nEq==pBuilder->nRecValid ){ 1764 UnpackedRecord *pRec = pBuilder->pRec; 1765 tRowcnt a[2]; 1766 int nBtm = pLoop->u.btree.nBtm; 1767 int nTop = pLoop->u.btree.nTop; 1768 1769 /* Variable iLower will be set to the estimate of the number of rows in 1770 ** the index that are less than the lower bound of the range query. The 1771 ** lower bound being the concatenation of $P and $L, where $P is the 1772 ** key-prefix formed by the nEq values matched against the nEq left-most 1773 ** columns of the index, and $L is the value in pLower. 1774 ** 1775 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1776 ** is not a simple variable or literal value), the lower bound of the 1777 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1778 ** if $L is available, whereKeyStats() is called for both ($P) and 1779 ** ($P:$L) and the larger of the two returned values is used. 1780 ** 1781 ** Similarly, iUpper is to be set to the estimate of the number of rows 1782 ** less than the upper bound of the range query. Where the upper bound 1783 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1784 ** of iUpper are requested of whereKeyStats() and the smaller used. 1785 ** 1786 ** The number of rows between the two bounds is then just iUpper-iLower. 1787 */ 1788 tRowcnt iLower; /* Rows less than the lower bound */ 1789 tRowcnt iUpper; /* Rows less than the upper bound */ 1790 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1791 int iUprIdx = -1; /* aSample[] for the upper bound */ 1792 1793 if( pRec ){ 1794 testcase( pRec->nField!=pBuilder->nRecValid ); 1795 pRec->nField = pBuilder->nRecValid; 1796 } 1797 /* Determine iLower and iUpper using ($P) only. */ 1798 if( nEq==0 ){ 1799 iLower = 0; 1800 iUpper = p->nRowEst0; 1801 }else{ 1802 /* Note: this call could be optimized away - since the same values must 1803 ** have been requested when testing key $P in whereEqualScanEst(). */ 1804 whereKeyStats(pParse, p, pRec, 0, a); 1805 iLower = a[0]; 1806 iUpper = a[0] + a[1]; 1807 } 1808 1809 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1810 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1811 assert( p->aSortOrder!=0 ); 1812 if( p->aSortOrder[nEq] ){ 1813 /* The roles of pLower and pUpper are swapped for a DESC index */ 1814 SWAP(WhereTerm*, pLower, pUpper); 1815 SWAP(int, nBtm, nTop); 1816 } 1817 1818 /* If possible, improve on the iLower estimate using ($P:$L). */ 1819 if( pLower ){ 1820 int n; /* Values extracted from pExpr */ 1821 Expr *pExpr = pLower->pExpr->pRight; 1822 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1823 if( rc==SQLITE_OK && n ){ 1824 tRowcnt iNew; 1825 u16 mask = WO_GT|WO_LE; 1826 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1827 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1828 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1829 if( iNew>iLower ) iLower = iNew; 1830 nOut--; 1831 pLower = 0; 1832 } 1833 } 1834 1835 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1836 if( pUpper ){ 1837 int n; /* Values extracted from pExpr */ 1838 Expr *pExpr = pUpper->pExpr->pRight; 1839 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1840 if( rc==SQLITE_OK && n ){ 1841 tRowcnt iNew; 1842 u16 mask = WO_GT|WO_LE; 1843 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1844 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1845 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1846 if( iNew<iUpper ) iUpper = iNew; 1847 nOut--; 1848 pUpper = 0; 1849 } 1850 } 1851 1852 pBuilder->pRec = pRec; 1853 if( rc==SQLITE_OK ){ 1854 if( iUpper>iLower ){ 1855 nNew = sqlite3LogEst(iUpper - iLower); 1856 /* TUNING: If both iUpper and iLower are derived from the same 1857 ** sample, then assume they are 4x more selective. This brings 1858 ** the estimated selectivity more in line with what it would be 1859 ** if estimated without the use of STAT4 tables. */ 1860 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1861 }else{ 1862 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1863 } 1864 if( nNew<nOut ){ 1865 nOut = nNew; 1866 } 1867 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1868 (u32)iLower, (u32)iUpper, nOut)); 1869 } 1870 }else{ 1871 int bDone = 0; 1872 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1873 if( bDone ) return rc; 1874 } 1875 } 1876 #else 1877 UNUSED_PARAMETER(pParse); 1878 UNUSED_PARAMETER(pBuilder); 1879 assert( pLower || pUpper ); 1880 #endif 1881 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1882 nNew = whereRangeAdjust(pLower, nOut); 1883 nNew = whereRangeAdjust(pUpper, nNew); 1884 1885 /* TUNING: If there is both an upper and lower limit and neither limit 1886 ** has an application-defined likelihood(), assume the range is 1887 ** reduced by an additional 75%. This means that, by default, an open-ended 1888 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1889 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1890 ** match 1/64 of the index. */ 1891 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1892 nNew -= 20; 1893 } 1894 1895 nOut -= (pLower!=0) + (pUpper!=0); 1896 if( nNew<10 ) nNew = 10; 1897 if( nNew<nOut ) nOut = nNew; 1898 #if defined(WHERETRACE_ENABLED) 1899 if( pLoop->nOut>nOut ){ 1900 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1901 pLoop->nOut, nOut)); 1902 } 1903 #endif 1904 pLoop->nOut = (LogEst)nOut; 1905 return rc; 1906 } 1907 1908 #ifdef SQLITE_ENABLE_STAT4 1909 /* 1910 ** Estimate the number of rows that will be returned based on 1911 ** an equality constraint x=VALUE and where that VALUE occurs in 1912 ** the histogram data. This only works when x is the left-most 1913 ** column of an index and sqlite_stat4 histogram data is available 1914 ** for that index. When pExpr==NULL that means the constraint is 1915 ** "x IS NULL" instead of "x=VALUE". 1916 ** 1917 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1918 ** If unable to make an estimate, leave *pnRow unchanged and return 1919 ** non-zero. 1920 ** 1921 ** This routine can fail if it is unable to load a collating sequence 1922 ** required for string comparison, or if unable to allocate memory 1923 ** for a UTF conversion required for comparison. The error is stored 1924 ** in the pParse structure. 1925 */ 1926 static int whereEqualScanEst( 1927 Parse *pParse, /* Parsing & code generating context */ 1928 WhereLoopBuilder *pBuilder, 1929 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1930 tRowcnt *pnRow /* Write the revised row estimate here */ 1931 ){ 1932 Index *p = pBuilder->pNew->u.btree.pIndex; 1933 int nEq = pBuilder->pNew->u.btree.nEq; 1934 UnpackedRecord *pRec = pBuilder->pRec; 1935 int rc; /* Subfunction return code */ 1936 tRowcnt a[2]; /* Statistics */ 1937 int bOk; 1938 1939 assert( nEq>=1 ); 1940 assert( nEq<=p->nColumn ); 1941 assert( p->aSample!=0 ); 1942 assert( p->nSample>0 ); 1943 assert( pBuilder->nRecValid<nEq ); 1944 1945 /* If values are not available for all fields of the index to the left 1946 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1947 if( pBuilder->nRecValid<(nEq-1) ){ 1948 return SQLITE_NOTFOUND; 1949 } 1950 1951 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1952 ** below would return the same value. */ 1953 if( nEq>=p->nColumn ){ 1954 *pnRow = 1; 1955 return SQLITE_OK; 1956 } 1957 1958 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1959 pBuilder->pRec = pRec; 1960 if( rc!=SQLITE_OK ) return rc; 1961 if( bOk==0 ) return SQLITE_NOTFOUND; 1962 pBuilder->nRecValid = nEq; 1963 1964 whereKeyStats(pParse, p, pRec, 0, a); 1965 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1966 p->zName, nEq-1, (int)a[1])); 1967 *pnRow = a[1]; 1968 1969 return rc; 1970 } 1971 #endif /* SQLITE_ENABLE_STAT4 */ 1972 1973 #ifdef SQLITE_ENABLE_STAT4 1974 /* 1975 ** Estimate the number of rows that will be returned based on 1976 ** an IN constraint where the right-hand side of the IN operator 1977 ** is a list of values. Example: 1978 ** 1979 ** WHERE x IN (1,2,3,4) 1980 ** 1981 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1982 ** If unable to make an estimate, leave *pnRow unchanged and return 1983 ** non-zero. 1984 ** 1985 ** This routine can fail if it is unable to load a collating sequence 1986 ** required for string comparison, or if unable to allocate memory 1987 ** for a UTF conversion required for comparison. The error is stored 1988 ** in the pParse structure. 1989 */ 1990 static int whereInScanEst( 1991 Parse *pParse, /* Parsing & code generating context */ 1992 WhereLoopBuilder *pBuilder, 1993 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1994 tRowcnt *pnRow /* Write the revised row estimate here */ 1995 ){ 1996 Index *p = pBuilder->pNew->u.btree.pIndex; 1997 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1998 int nRecValid = pBuilder->nRecValid; 1999 int rc = SQLITE_OK; /* Subfunction return code */ 2000 tRowcnt nEst; /* Number of rows for a single term */ 2001 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 2002 int i; /* Loop counter */ 2003 2004 assert( p->aSample!=0 ); 2005 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 2006 nEst = nRow0; 2007 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 2008 nRowEst += nEst; 2009 pBuilder->nRecValid = nRecValid; 2010 } 2011 2012 if( rc==SQLITE_OK ){ 2013 if( nRowEst > nRow0 ) nRowEst = nRow0; 2014 *pnRow = nRowEst; 2015 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 2016 } 2017 assert( pBuilder->nRecValid==nRecValid ); 2018 return rc; 2019 } 2020 #endif /* SQLITE_ENABLE_STAT4 */ 2021 2022 2023 #ifdef WHERETRACE_ENABLED 2024 /* 2025 ** Print the content of a WhereTerm object 2026 */ 2027 void sqlite3WhereTermPrint(WhereTerm *pTerm, int iTerm){ 2028 if( pTerm==0 ){ 2029 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 2030 }else{ 2031 char zType[8]; 2032 char zLeft[50]; 2033 memcpy(zType, "....", 5); 2034 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 2035 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 2036 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 2037 if( pTerm->wtFlags & TERM_CODED ) zType[3] = 'C'; 2038 if( pTerm->eOperator & WO_SINGLE ){ 2039 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 ); 2040 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 2041 pTerm->leftCursor, pTerm->u.x.leftColumn); 2042 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 2043 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%llx", 2044 pTerm->u.pOrInfo->indexable); 2045 }else{ 2046 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 2047 } 2048 sqlite3DebugPrintf( 2049 "TERM-%-3d %p %s %-12s op=%03x wtFlags=%04x", 2050 iTerm, pTerm, zType, zLeft, pTerm->eOperator, pTerm->wtFlags); 2051 /* The 0x10000 .wheretrace flag causes extra information to be 2052 ** shown about each Term */ 2053 if( sqlite3WhereTrace & 0x10000 ){ 2054 sqlite3DebugPrintf(" prob=%-3d prereq=%llx,%llx", 2055 pTerm->truthProb, (u64)pTerm->prereqAll, (u64)pTerm->prereqRight); 2056 } 2057 if( (pTerm->eOperator & (WO_OR|WO_AND))==0 && pTerm->u.x.iField ){ 2058 sqlite3DebugPrintf(" iField=%d", pTerm->u.x.iField); 2059 } 2060 if( pTerm->iParent>=0 ){ 2061 sqlite3DebugPrintf(" iParent=%d", pTerm->iParent); 2062 } 2063 sqlite3DebugPrintf("\n"); 2064 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 2065 } 2066 } 2067 #endif 2068 2069 #ifdef WHERETRACE_ENABLED 2070 /* 2071 ** Show the complete content of a WhereClause 2072 */ 2073 void sqlite3WhereClausePrint(WhereClause *pWC){ 2074 int i; 2075 for(i=0; i<pWC->nTerm; i++){ 2076 sqlite3WhereTermPrint(&pWC->a[i], i); 2077 } 2078 } 2079 #endif 2080 2081 #ifdef WHERETRACE_ENABLED 2082 /* 2083 ** Print a WhereLoop object for debugging purposes 2084 */ 2085 void sqlite3WhereLoopPrint(WhereLoop *p, WhereClause *pWC){ 2086 WhereInfo *pWInfo = pWC->pWInfo; 2087 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 2088 SrcItem *pItem = pWInfo->pTabList->a + p->iTab; 2089 Table *pTab = pItem->pTab; 2090 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 2091 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 2092 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 2093 sqlite3DebugPrintf(" %12s", 2094 pItem->zAlias ? pItem->zAlias : pTab->zName); 2095 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2096 const char *zName; 2097 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 2098 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 2099 int i = sqlite3Strlen30(zName) - 1; 2100 while( zName[i]!='_' ) i--; 2101 zName += i; 2102 } 2103 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 2104 }else{ 2105 sqlite3DebugPrintf("%20s",""); 2106 } 2107 }else{ 2108 char *z; 2109 if( p->u.vtab.idxStr ){ 2110 z = sqlite3_mprintf("(%d,\"%s\",%#x)", 2111 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 2112 }else{ 2113 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 2114 } 2115 sqlite3DebugPrintf(" %-19s", z); 2116 sqlite3_free(z); 2117 } 2118 if( p->wsFlags & WHERE_SKIPSCAN ){ 2119 sqlite3DebugPrintf(" f %06x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 2120 }else{ 2121 sqlite3DebugPrintf(" f %06x N %d", p->wsFlags, p->nLTerm); 2122 } 2123 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 2124 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 2125 int i; 2126 for(i=0; i<p->nLTerm; i++){ 2127 sqlite3WhereTermPrint(p->aLTerm[i], i); 2128 } 2129 } 2130 } 2131 #endif 2132 2133 /* 2134 ** Convert bulk memory into a valid WhereLoop that can be passed 2135 ** to whereLoopClear harmlessly. 2136 */ 2137 static void whereLoopInit(WhereLoop *p){ 2138 p->aLTerm = p->aLTermSpace; 2139 p->nLTerm = 0; 2140 p->nLSlot = ArraySize(p->aLTermSpace); 2141 p->wsFlags = 0; 2142 } 2143 2144 /* 2145 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 2146 */ 2147 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 2148 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 2149 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 2150 sqlite3_free(p->u.vtab.idxStr); 2151 p->u.vtab.needFree = 0; 2152 p->u.vtab.idxStr = 0; 2153 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 2154 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 2155 sqlite3DbFreeNN(db, p->u.btree.pIndex); 2156 p->u.btree.pIndex = 0; 2157 } 2158 } 2159 } 2160 2161 /* 2162 ** Deallocate internal memory used by a WhereLoop object 2163 */ 2164 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 2165 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 2166 whereLoopClearUnion(db, p); 2167 whereLoopInit(p); 2168 } 2169 2170 /* 2171 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 2172 */ 2173 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 2174 WhereTerm **paNew; 2175 if( p->nLSlot>=n ) return SQLITE_OK; 2176 n = (n+7)&~7; 2177 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 2178 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 2179 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 2180 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 2181 p->aLTerm = paNew; 2182 p->nLSlot = n; 2183 return SQLITE_OK; 2184 } 2185 2186 /* 2187 ** Transfer content from the second pLoop into the first. 2188 */ 2189 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 2190 whereLoopClearUnion(db, pTo); 2191 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 2192 memset(pTo, 0, WHERE_LOOP_XFER_SZ); 2193 return SQLITE_NOMEM_BKPT; 2194 } 2195 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 2196 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 2197 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 2198 pFrom->u.vtab.needFree = 0; 2199 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 2200 pFrom->u.btree.pIndex = 0; 2201 } 2202 return SQLITE_OK; 2203 } 2204 2205 /* 2206 ** Delete a WhereLoop object 2207 */ 2208 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 2209 whereLoopClear(db, p); 2210 sqlite3DbFreeNN(db, p); 2211 } 2212 2213 /* 2214 ** Free a WhereInfo structure 2215 */ 2216 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 2217 int i; 2218 assert( pWInfo!=0 ); 2219 for(i=0; i<pWInfo->nLevel; i++){ 2220 WhereLevel *pLevel = &pWInfo->a[i]; 2221 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE)!=0 ){ 2222 assert( (pLevel->pWLoop->wsFlags & WHERE_MULTI_OR)==0 ); 2223 sqlite3DbFree(db, pLevel->u.in.aInLoop); 2224 } 2225 } 2226 sqlite3WhereClauseClear(&pWInfo->sWC); 2227 while( pWInfo->pLoops ){ 2228 WhereLoop *p = pWInfo->pLoops; 2229 pWInfo->pLoops = p->pNextLoop; 2230 whereLoopDelete(db, p); 2231 } 2232 assert( pWInfo->pExprMods==0 ); 2233 sqlite3DbFreeNN(db, pWInfo); 2234 } 2235 2236 /* Undo all Expr node modifications 2237 */ 2238 static void whereUndoExprMods(WhereInfo *pWInfo){ 2239 while( pWInfo->pExprMods ){ 2240 WhereExprMod *p = pWInfo->pExprMods; 2241 pWInfo->pExprMods = p->pNext; 2242 memcpy(p->pExpr, &p->orig, sizeof(p->orig)); 2243 sqlite3DbFree(pWInfo->pParse->db, p); 2244 } 2245 } 2246 2247 /* 2248 ** Return TRUE if all of the following are true: 2249 ** 2250 ** (1) X has the same or lower cost, or returns the same or fewer rows, 2251 ** than Y. 2252 ** (2) X uses fewer WHERE clause terms than Y 2253 ** (3) Every WHERE clause term used by X is also used by Y 2254 ** (4) X skips at least as many columns as Y 2255 ** (5) If X is a covering index, than Y is too 2256 ** 2257 ** Conditions (2) and (3) mean that X is a "proper subset" of Y. 2258 ** If X is a proper subset of Y then Y is a better choice and ought 2259 ** to have a lower cost. This routine returns TRUE when that cost 2260 ** relationship is inverted and needs to be adjusted. Constraint (4) 2261 ** was added because if X uses skip-scan less than Y it still might 2262 ** deserve a lower cost even if it is a proper subset of Y. Constraint (5) 2263 ** was added because a covering index probably deserves to have a lower cost 2264 ** than a non-covering index even if it is a proper subset. 2265 */ 2266 static int whereLoopCheaperProperSubset( 2267 const WhereLoop *pX, /* First WhereLoop to compare */ 2268 const WhereLoop *pY /* Compare against this WhereLoop */ 2269 ){ 2270 int i, j; 2271 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 2272 return 0; /* X is not a subset of Y */ 2273 } 2274 if( pX->rRun>pY->rRun && pX->nOut>pY->nOut ) return 0; 2275 if( pY->nSkip > pX->nSkip ) return 0; 2276 for(i=pX->nLTerm-1; i>=0; i--){ 2277 if( pX->aLTerm[i]==0 ) continue; 2278 for(j=pY->nLTerm-1; j>=0; j--){ 2279 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 2280 } 2281 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 2282 } 2283 if( (pX->wsFlags&WHERE_IDX_ONLY)!=0 2284 && (pY->wsFlags&WHERE_IDX_ONLY)==0 ){ 2285 return 0; /* Constraint (5) */ 2286 } 2287 return 1; /* All conditions meet */ 2288 } 2289 2290 /* 2291 ** Try to adjust the cost and number of output rows of WhereLoop pTemplate 2292 ** upwards or downwards so that: 2293 ** 2294 ** (1) pTemplate costs less than any other WhereLoops that are a proper 2295 ** subset of pTemplate 2296 ** 2297 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 2298 ** is a proper subset. 2299 ** 2300 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 2301 ** WHERE clause terms than Y and that every WHERE clause term used by X is 2302 ** also used by Y. 2303 */ 2304 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 2305 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 2306 for(; p; p=p->pNextLoop){ 2307 if( p->iTab!=pTemplate->iTab ) continue; 2308 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 2309 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 2310 /* Adjust pTemplate cost downward so that it is cheaper than its 2311 ** subset p. */ 2312 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2313 pTemplate->rRun, pTemplate->nOut, 2314 MIN(p->rRun, pTemplate->rRun), 2315 MIN(p->nOut - 1, pTemplate->nOut))); 2316 pTemplate->rRun = MIN(p->rRun, pTemplate->rRun); 2317 pTemplate->nOut = MIN(p->nOut - 1, pTemplate->nOut); 2318 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 2319 /* Adjust pTemplate cost upward so that it is costlier than p since 2320 ** pTemplate is a proper subset of p */ 2321 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2322 pTemplate->rRun, pTemplate->nOut, 2323 MAX(p->rRun, pTemplate->rRun), 2324 MAX(p->nOut + 1, pTemplate->nOut))); 2325 pTemplate->rRun = MAX(p->rRun, pTemplate->rRun); 2326 pTemplate->nOut = MAX(p->nOut + 1, pTemplate->nOut); 2327 } 2328 } 2329 } 2330 2331 /* 2332 ** Search the list of WhereLoops in *ppPrev looking for one that can be 2333 ** replaced by pTemplate. 2334 ** 2335 ** Return NULL if pTemplate does not belong on the WhereLoop list. 2336 ** In other words if pTemplate ought to be dropped from further consideration. 2337 ** 2338 ** If pX is a WhereLoop that pTemplate can replace, then return the 2339 ** link that points to pX. 2340 ** 2341 ** If pTemplate cannot replace any existing element of the list but needs 2342 ** to be added to the list as a new entry, then return a pointer to the 2343 ** tail of the list. 2344 */ 2345 static WhereLoop **whereLoopFindLesser( 2346 WhereLoop **ppPrev, 2347 const WhereLoop *pTemplate 2348 ){ 2349 WhereLoop *p; 2350 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 2351 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 2352 /* If either the iTab or iSortIdx values for two WhereLoop are different 2353 ** then those WhereLoops need to be considered separately. Neither is 2354 ** a candidate to replace the other. */ 2355 continue; 2356 } 2357 /* In the current implementation, the rSetup value is either zero 2358 ** or the cost of building an automatic index (NlogN) and the NlogN 2359 ** is the same for compatible WhereLoops. */ 2360 assert( p->rSetup==0 || pTemplate->rSetup==0 2361 || p->rSetup==pTemplate->rSetup ); 2362 2363 /* whereLoopAddBtree() always generates and inserts the automatic index 2364 ** case first. Hence compatible candidate WhereLoops never have a larger 2365 ** rSetup. Call this SETUP-INVARIANT */ 2366 assert( p->rSetup>=pTemplate->rSetup ); 2367 2368 /* Any loop using an appliation-defined index (or PRIMARY KEY or 2369 ** UNIQUE constraint) with one or more == constraints is better 2370 ** than an automatic index. Unless it is a skip-scan. */ 2371 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 2372 && (pTemplate->nSkip)==0 2373 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 2374 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 2375 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2376 ){ 2377 break; 2378 } 2379 2380 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2381 ** discarded. WhereLoop p is better if: 2382 ** (1) p has no more dependencies than pTemplate, and 2383 ** (2) p has an equal or lower cost than pTemplate 2384 */ 2385 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2386 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2387 && p->rRun<=pTemplate->rRun /* (2b) */ 2388 && p->nOut<=pTemplate->nOut /* (2c) */ 2389 ){ 2390 return 0; /* Discard pTemplate */ 2391 } 2392 2393 /* If pTemplate is always better than p, then cause p to be overwritten 2394 ** with pTemplate. pTemplate is better than p if: 2395 ** (1) pTemplate has no more dependences than p, and 2396 ** (2) pTemplate has an equal or lower cost than p. 2397 */ 2398 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2399 && p->rRun>=pTemplate->rRun /* (2a) */ 2400 && p->nOut>=pTemplate->nOut /* (2b) */ 2401 ){ 2402 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2403 break; /* Cause p to be overwritten by pTemplate */ 2404 } 2405 } 2406 return ppPrev; 2407 } 2408 2409 /* 2410 ** Insert or replace a WhereLoop entry using the template supplied. 2411 ** 2412 ** An existing WhereLoop entry might be overwritten if the new template 2413 ** is better and has fewer dependencies. Or the template will be ignored 2414 ** and no insert will occur if an existing WhereLoop is faster and has 2415 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2416 ** added based on the template. 2417 ** 2418 ** If pBuilder->pOrSet is not NULL then we care about only the 2419 ** prerequisites and rRun and nOut costs of the N best loops. That 2420 ** information is gathered in the pBuilder->pOrSet object. This special 2421 ** processing mode is used only for OR clause processing. 2422 ** 2423 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2424 ** still might overwrite similar loops with the new template if the 2425 ** new template is better. Loops may be overwritten if the following 2426 ** conditions are met: 2427 ** 2428 ** (1) They have the same iTab. 2429 ** (2) They have the same iSortIdx. 2430 ** (3) The template has same or fewer dependencies than the current loop 2431 ** (4) The template has the same or lower cost than the current loop 2432 */ 2433 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2434 WhereLoop **ppPrev, *p; 2435 WhereInfo *pWInfo = pBuilder->pWInfo; 2436 sqlite3 *db = pWInfo->pParse->db; 2437 int rc; 2438 2439 /* Stop the search once we hit the query planner search limit */ 2440 if( pBuilder->iPlanLimit==0 ){ 2441 WHERETRACE(0xffffffff,("=== query planner search limit reached ===\n")); 2442 if( pBuilder->pOrSet ) pBuilder->pOrSet->n = 0; 2443 return SQLITE_DONE; 2444 } 2445 pBuilder->iPlanLimit--; 2446 2447 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2448 2449 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2450 ** and prereqs. 2451 */ 2452 if( pBuilder->pOrSet!=0 ){ 2453 if( pTemplate->nLTerm ){ 2454 #if WHERETRACE_ENABLED 2455 u16 n = pBuilder->pOrSet->n; 2456 int x = 2457 #endif 2458 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2459 pTemplate->nOut); 2460 #if WHERETRACE_ENABLED /* 0x8 */ 2461 if( sqlite3WhereTrace & 0x8 ){ 2462 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2463 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2464 } 2465 #endif 2466 } 2467 return SQLITE_OK; 2468 } 2469 2470 /* Look for an existing WhereLoop to replace with pTemplate 2471 */ 2472 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2473 2474 if( ppPrev==0 ){ 2475 /* There already exists a WhereLoop on the list that is better 2476 ** than pTemplate, so just ignore pTemplate */ 2477 #if WHERETRACE_ENABLED /* 0x8 */ 2478 if( sqlite3WhereTrace & 0x8 ){ 2479 sqlite3DebugPrintf(" skip: "); 2480 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2481 } 2482 #endif 2483 return SQLITE_OK; 2484 }else{ 2485 p = *ppPrev; 2486 } 2487 2488 /* If we reach this point it means that either p[] should be overwritten 2489 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2490 ** WhereLoop and insert it. 2491 */ 2492 #if WHERETRACE_ENABLED /* 0x8 */ 2493 if( sqlite3WhereTrace & 0x8 ){ 2494 if( p!=0 ){ 2495 sqlite3DebugPrintf("replace: "); 2496 sqlite3WhereLoopPrint(p, pBuilder->pWC); 2497 sqlite3DebugPrintf(" with: "); 2498 }else{ 2499 sqlite3DebugPrintf(" add: "); 2500 } 2501 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2502 } 2503 #endif 2504 if( p==0 ){ 2505 /* Allocate a new WhereLoop to add to the end of the list */ 2506 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2507 if( p==0 ) return SQLITE_NOMEM_BKPT; 2508 whereLoopInit(p); 2509 p->pNextLoop = 0; 2510 }else{ 2511 /* We will be overwriting WhereLoop p[]. But before we do, first 2512 ** go through the rest of the list and delete any other entries besides 2513 ** p[] that are also supplated by pTemplate */ 2514 WhereLoop **ppTail = &p->pNextLoop; 2515 WhereLoop *pToDel; 2516 while( *ppTail ){ 2517 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2518 if( ppTail==0 ) break; 2519 pToDel = *ppTail; 2520 if( pToDel==0 ) break; 2521 *ppTail = pToDel->pNextLoop; 2522 #if WHERETRACE_ENABLED /* 0x8 */ 2523 if( sqlite3WhereTrace & 0x8 ){ 2524 sqlite3DebugPrintf(" delete: "); 2525 sqlite3WhereLoopPrint(pToDel, pBuilder->pWC); 2526 } 2527 #endif 2528 whereLoopDelete(db, pToDel); 2529 } 2530 } 2531 rc = whereLoopXfer(db, p, pTemplate); 2532 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2533 Index *pIndex = p->u.btree.pIndex; 2534 if( pIndex && pIndex->idxType==SQLITE_IDXTYPE_IPK ){ 2535 p->u.btree.pIndex = 0; 2536 } 2537 } 2538 return rc; 2539 } 2540 2541 /* 2542 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2543 ** WHERE clause that reference the loop but which are not used by an 2544 ** index. 2545 * 2546 ** For every WHERE clause term that is not used by the index 2547 ** and which has a truth probability assigned by one of the likelihood(), 2548 ** likely(), or unlikely() SQL functions, reduce the estimated number 2549 ** of output rows by the probability specified. 2550 ** 2551 ** TUNING: For every WHERE clause term that is not used by the index 2552 ** and which does not have an assigned truth probability, heuristics 2553 ** described below are used to try to estimate the truth probability. 2554 ** TODO --> Perhaps this is something that could be improved by better 2555 ** table statistics. 2556 ** 2557 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2558 ** value corresponds to -1 in LogEst notation, so this means decrement 2559 ** the WhereLoop.nOut field for every such WHERE clause term. 2560 ** 2561 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2562 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2563 ** final output row estimate is no greater than 1/4 of the total number 2564 ** of rows in the table. In other words, assume that x==EXPR will filter 2565 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2566 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2567 ** on the "x" column and so in that case only cap the output row estimate 2568 ** at 1/2 instead of 1/4. 2569 */ 2570 static void whereLoopOutputAdjust( 2571 WhereClause *pWC, /* The WHERE clause */ 2572 WhereLoop *pLoop, /* The loop to adjust downward */ 2573 LogEst nRow /* Number of rows in the entire table */ 2574 ){ 2575 WhereTerm *pTerm, *pX; 2576 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2577 int i, j; 2578 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2579 2580 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2581 for(i=pWC->nBase, pTerm=pWC->a; i>0; i--, pTerm++){ 2582 assert( pTerm!=0 ); 2583 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2584 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2585 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) continue; 2586 for(j=pLoop->nLTerm-1; j>=0; j--){ 2587 pX = pLoop->aLTerm[j]; 2588 if( pX==0 ) continue; 2589 if( pX==pTerm ) break; 2590 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2591 } 2592 if( j<0 ){ 2593 if( pLoop->maskSelf==pTerm->prereqAll ){ 2594 /* If there are extra terms in the WHERE clause not used by an index 2595 ** that depend only on the table being scanned, and that will tend to 2596 ** cause many rows to be omitted, then mark that table as 2597 ** "self-culling". */ 2598 pLoop->wsFlags |= WHERE_SELFCULL; 2599 } 2600 if( pTerm->truthProb<=0 ){ 2601 /* If a truth probability is specified using the likelihood() hints, 2602 ** then use the probability provided by the application. */ 2603 pLoop->nOut += pTerm->truthProb; 2604 }else{ 2605 /* In the absence of explicit truth probabilities, use heuristics to 2606 ** guess a reasonable truth probability. */ 2607 pLoop->nOut--; 2608 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 2609 && (pTerm->wtFlags & TERM_HIGHTRUTH)==0 /* tag-20200224-1 */ 2610 ){ 2611 Expr *pRight = pTerm->pExpr->pRight; 2612 int k = 0; 2613 testcase( pTerm->pExpr->op==TK_IS ); 2614 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2615 k = 10; 2616 }else{ 2617 k = 20; 2618 } 2619 if( iReduce<k ){ 2620 pTerm->wtFlags |= TERM_HEURTRUTH; 2621 iReduce = k; 2622 } 2623 } 2624 } 2625 } 2626 } 2627 if( pLoop->nOut > nRow-iReduce ){ 2628 pLoop->nOut = nRow - iReduce; 2629 } 2630 } 2631 2632 /* 2633 ** Term pTerm is a vector range comparison operation. The first comparison 2634 ** in the vector can be optimized using column nEq of the index. This 2635 ** function returns the total number of vector elements that can be used 2636 ** as part of the range comparison. 2637 ** 2638 ** For example, if the query is: 2639 ** 2640 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2641 ** 2642 ** and the index: 2643 ** 2644 ** CREATE INDEX ... ON (a, b, c, d, e) 2645 ** 2646 ** then this function would be invoked with nEq=1. The value returned in 2647 ** this case is 3. 2648 */ 2649 static int whereRangeVectorLen( 2650 Parse *pParse, /* Parsing context */ 2651 int iCur, /* Cursor open on pIdx */ 2652 Index *pIdx, /* The index to be used for a inequality constraint */ 2653 int nEq, /* Number of prior equality constraints on same index */ 2654 WhereTerm *pTerm /* The vector inequality constraint */ 2655 ){ 2656 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2657 int i; 2658 2659 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2660 for(i=1; i<nCmp; i++){ 2661 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2662 ** of the index. If not, exit the loop. */ 2663 char aff; /* Comparison affinity */ 2664 char idxaff = 0; /* Indexed columns affinity */ 2665 CollSeq *pColl; /* Comparison collation sequence */ 2666 Expr *pLhs, *pRhs; 2667 2668 assert( ExprUseXList(pTerm->pExpr->pLeft) ); 2669 pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2670 pRhs = pTerm->pExpr->pRight; 2671 if( ExprUseXSelect(pRhs) ){ 2672 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2673 }else{ 2674 pRhs = pRhs->x.pList->a[i].pExpr; 2675 } 2676 2677 /* Check that the LHS of the comparison is a column reference to 2678 ** the right column of the right source table. And that the sort 2679 ** order of the index column is the same as the sort order of the 2680 ** leftmost index column. */ 2681 if( pLhs->op!=TK_COLUMN 2682 || pLhs->iTable!=iCur 2683 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2684 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2685 ){ 2686 break; 2687 } 2688 2689 testcase( pLhs->iColumn==XN_ROWID ); 2690 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2691 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2692 if( aff!=idxaff ) break; 2693 2694 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2695 if( pColl==0 ) break; 2696 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2697 } 2698 return i; 2699 } 2700 2701 /* 2702 ** Adjust the cost C by the costMult facter T. This only occurs if 2703 ** compiled with -DSQLITE_ENABLE_COSTMULT 2704 */ 2705 #ifdef SQLITE_ENABLE_COSTMULT 2706 # define ApplyCostMultiplier(C,T) C += T 2707 #else 2708 # define ApplyCostMultiplier(C,T) 2709 #endif 2710 2711 /* 2712 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2713 ** index pIndex. Try to match one more. 2714 ** 2715 ** When this function is called, pBuilder->pNew->nOut contains the 2716 ** number of rows expected to be visited by filtering using the nEq 2717 ** terms only. If it is modified, this value is restored before this 2718 ** function returns. 2719 ** 2720 ** If pProbe->idxType==SQLITE_IDXTYPE_IPK, that means pIndex is 2721 ** a fake index used for the INTEGER PRIMARY KEY. 2722 */ 2723 static int whereLoopAddBtreeIndex( 2724 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2725 SrcItem *pSrc, /* FROM clause term being analyzed */ 2726 Index *pProbe, /* An index on pSrc */ 2727 LogEst nInMul /* log(Number of iterations due to IN) */ 2728 ){ 2729 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2730 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2731 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2732 WhereLoop *pNew; /* Template WhereLoop under construction */ 2733 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2734 int opMask; /* Valid operators for constraints */ 2735 WhereScan scan; /* Iterator for WHERE terms */ 2736 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2737 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2738 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2739 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2740 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2741 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2742 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2743 LogEst saved_nOut; /* Original value of pNew->nOut */ 2744 int rc = SQLITE_OK; /* Return code */ 2745 LogEst rSize; /* Number of rows in the table */ 2746 LogEst rLogSize; /* Logarithm of table size */ 2747 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2748 2749 pNew = pBuilder->pNew; 2750 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2751 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d, nSkip=%d, rRun=%d\n", 2752 pProbe->pTable->zName,pProbe->zName, 2753 pNew->u.btree.nEq, pNew->nSkip, pNew->rRun)); 2754 2755 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2756 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2757 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2758 opMask = WO_LT|WO_LE; 2759 }else{ 2760 assert( pNew->u.btree.nBtm==0 ); 2761 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2762 } 2763 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2764 2765 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2766 assert( pNew->u.btree.nEq<pProbe->nKeyCol 2767 || pProbe->idxType!=SQLITE_IDXTYPE_PRIMARYKEY ); 2768 2769 saved_nEq = pNew->u.btree.nEq; 2770 saved_nBtm = pNew->u.btree.nBtm; 2771 saved_nTop = pNew->u.btree.nTop; 2772 saved_nSkip = pNew->nSkip; 2773 saved_nLTerm = pNew->nLTerm; 2774 saved_wsFlags = pNew->wsFlags; 2775 saved_prereq = pNew->prereq; 2776 saved_nOut = pNew->nOut; 2777 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2778 opMask, pProbe); 2779 pNew->rSetup = 0; 2780 rSize = pProbe->aiRowLogEst[0]; 2781 rLogSize = estLog(rSize); 2782 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2783 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2784 LogEst rCostIdx; 2785 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2786 int nIn = 0; 2787 #ifdef SQLITE_ENABLE_STAT4 2788 int nRecValid = pBuilder->nRecValid; 2789 #endif 2790 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2791 && indexColumnNotNull(pProbe, saved_nEq) 2792 ){ 2793 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2794 } 2795 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2796 2797 /* Do not allow the upper bound of a LIKE optimization range constraint 2798 ** to mix with a lower range bound from some other source */ 2799 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2800 2801 /* tag-20191211-001: Do not allow constraints from the WHERE clause to 2802 ** be used by the right table of a LEFT JOIN. Only constraints in the 2803 ** ON clause are allowed. See tag-20191211-002 for the vtab equivalent. */ 2804 if( (pSrc->fg.jointype & JT_LEFT)!=0 2805 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2806 ){ 2807 continue; 2808 } 2809 2810 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2811 pBuilder->bldFlags1 |= SQLITE_BLDF1_UNIQUE; 2812 }else{ 2813 pBuilder->bldFlags1 |= SQLITE_BLDF1_INDEXED; 2814 } 2815 pNew->wsFlags = saved_wsFlags; 2816 pNew->u.btree.nEq = saved_nEq; 2817 pNew->u.btree.nBtm = saved_nBtm; 2818 pNew->u.btree.nTop = saved_nTop; 2819 pNew->nLTerm = saved_nLTerm; 2820 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2821 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2822 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2823 2824 assert( nInMul==0 2825 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2826 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2827 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2828 ); 2829 2830 if( eOp & WO_IN ){ 2831 Expr *pExpr = pTerm->pExpr; 2832 if( ExprUseXSelect(pExpr) ){ 2833 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2834 int i; 2835 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2836 2837 /* The expression may actually be of the form (x, y) IN (SELECT...). 2838 ** In this case there is a separate term for each of (x) and (y). 2839 ** However, the nIn multiplier should only be applied once, not once 2840 ** for each such term. The following loop checks that pTerm is the 2841 ** first such term in use, and sets nIn back to 0 if it is not. */ 2842 for(i=0; i<pNew->nLTerm-1; i++){ 2843 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2844 } 2845 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2846 /* "x IN (value, value, ...)" */ 2847 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2848 } 2849 if( pProbe->hasStat1 && rLogSize>=10 ){ 2850 LogEst M, logK, x; 2851 /* Let: 2852 ** N = the total number of rows in the table 2853 ** K = the number of entries on the RHS of the IN operator 2854 ** M = the number of rows in the table that match terms to the 2855 ** to the left in the same index. If the IN operator is on 2856 ** the left-most index column, M==N. 2857 ** 2858 ** Given the definitions above, it is better to omit the IN operator 2859 ** from the index lookup and instead do a scan of the M elements, 2860 ** testing each scanned row against the IN operator separately, if: 2861 ** 2862 ** M*log(K) < K*log(N) 2863 ** 2864 ** Our estimates for M, K, and N might be inaccurate, so we build in 2865 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator 2866 ** with the index, as using an index has better worst-case behavior. 2867 ** If we do not have real sqlite_stat1 data, always prefer to use 2868 ** the index. Do not bother with this optimization on very small 2869 ** tables (less than 2 rows) as it is pointless in that case. 2870 */ 2871 M = pProbe->aiRowLogEst[saved_nEq]; 2872 logK = estLog(nIn); 2873 /* TUNING v----- 10 to bias toward indexed IN */ 2874 x = M + logK + 10 - (nIn + rLogSize); 2875 if( x>=0 ){ 2876 WHERETRACE(0x40, 2877 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d) " 2878 "prefers indexed lookup\n", 2879 saved_nEq, M, logK, nIn, rLogSize, x)); 2880 }else if( nInMul<2 && OptimizationEnabled(db, SQLITE_SeekScan) ){ 2881 WHERETRACE(0x40, 2882 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d" 2883 " nInMul=%d) prefers skip-scan\n", 2884 saved_nEq, M, logK, nIn, rLogSize, x, nInMul)); 2885 pNew->wsFlags |= WHERE_IN_SEEKSCAN; 2886 }else{ 2887 WHERETRACE(0x40, 2888 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d" 2889 " nInMul=%d) prefers normal scan\n", 2890 saved_nEq, M, logK, nIn, rLogSize, x, nInMul)); 2891 continue; 2892 } 2893 } 2894 pNew->wsFlags |= WHERE_COLUMN_IN; 2895 }else if( eOp & (WO_EQ|WO_IS) ){ 2896 int iCol = pProbe->aiColumn[saved_nEq]; 2897 pNew->wsFlags |= WHERE_COLUMN_EQ; 2898 assert( saved_nEq==pNew->u.btree.nEq ); 2899 if( iCol==XN_ROWID 2900 || (iCol>=0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2901 ){ 2902 if( iCol==XN_ROWID || pProbe->uniqNotNull 2903 || (pProbe->nKeyCol==1 && pProbe->onError && eOp==WO_EQ) 2904 ){ 2905 pNew->wsFlags |= WHERE_ONEROW; 2906 }else{ 2907 pNew->wsFlags |= WHERE_UNQ_WANTED; 2908 } 2909 } 2910 if( scan.iEquiv>1 ) pNew->wsFlags |= WHERE_TRANSCONS; 2911 }else if( eOp & WO_ISNULL ){ 2912 pNew->wsFlags |= WHERE_COLUMN_NULL; 2913 }else if( eOp & (WO_GT|WO_GE) ){ 2914 testcase( eOp & WO_GT ); 2915 testcase( eOp & WO_GE ); 2916 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2917 pNew->u.btree.nBtm = whereRangeVectorLen( 2918 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2919 ); 2920 pBtm = pTerm; 2921 pTop = 0; 2922 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2923 /* Range constraints that come from the LIKE optimization are 2924 ** always used in pairs. */ 2925 pTop = &pTerm[1]; 2926 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2927 assert( pTop->wtFlags & TERM_LIKEOPT ); 2928 assert( pTop->eOperator==WO_LT ); 2929 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2930 pNew->aLTerm[pNew->nLTerm++] = pTop; 2931 pNew->wsFlags |= WHERE_TOP_LIMIT; 2932 pNew->u.btree.nTop = 1; 2933 } 2934 }else{ 2935 assert( eOp & (WO_LT|WO_LE) ); 2936 testcase( eOp & WO_LT ); 2937 testcase( eOp & WO_LE ); 2938 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2939 pNew->u.btree.nTop = whereRangeVectorLen( 2940 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2941 ); 2942 pTop = pTerm; 2943 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2944 pNew->aLTerm[pNew->nLTerm-2] : 0; 2945 } 2946 2947 /* At this point pNew->nOut is set to the number of rows expected to 2948 ** be visited by the index scan before considering term pTerm, or the 2949 ** values of nIn and nInMul. In other words, assuming that all 2950 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2951 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2952 assert( pNew->nOut==saved_nOut ); 2953 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2954 /* Adjust nOut using stat4 data. Or, if there is no stat4 2955 ** data, using some other estimate. */ 2956 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2957 }else{ 2958 int nEq = ++pNew->u.btree.nEq; 2959 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2960 2961 assert( pNew->nOut==saved_nOut ); 2962 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2963 assert( (eOp & WO_IN) || nIn==0 ); 2964 testcase( eOp & WO_IN ); 2965 pNew->nOut += pTerm->truthProb; 2966 pNew->nOut -= nIn; 2967 }else{ 2968 #ifdef SQLITE_ENABLE_STAT4 2969 tRowcnt nOut = 0; 2970 if( nInMul==0 2971 && pProbe->nSample 2972 && ALWAYS(pNew->u.btree.nEq<=pProbe->nSampleCol) 2973 && ((eOp & WO_IN)==0 || ExprUseXList(pTerm->pExpr)) 2974 && OptimizationEnabled(db, SQLITE_Stat4) 2975 ){ 2976 Expr *pExpr = pTerm->pExpr; 2977 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2978 testcase( eOp & WO_EQ ); 2979 testcase( eOp & WO_IS ); 2980 testcase( eOp & WO_ISNULL ); 2981 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2982 }else{ 2983 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2984 } 2985 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2986 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2987 if( nOut ){ 2988 pNew->nOut = sqlite3LogEst(nOut); 2989 if( nEq==1 2990 /* TUNING: Mark terms as "low selectivity" if they seem likely 2991 ** to be true for half or more of the rows in the table. 2992 ** See tag-202002240-1 */ 2993 && pNew->nOut+10 > pProbe->aiRowLogEst[0] 2994 ){ 2995 #if WHERETRACE_ENABLED /* 0x01 */ 2996 if( sqlite3WhereTrace & 0x01 ){ 2997 sqlite3DebugPrintf( 2998 "STAT4 determines term has low selectivity:\n"); 2999 sqlite3WhereTermPrint(pTerm, 999); 3000 } 3001 #endif 3002 pTerm->wtFlags |= TERM_HIGHTRUTH; 3003 if( pTerm->wtFlags & TERM_HEURTRUTH ){ 3004 /* If the term has previously been used with an assumption of 3005 ** higher selectivity, then set the flag to rerun the 3006 ** loop computations. */ 3007 pBuilder->bldFlags2 |= SQLITE_BLDF2_2NDPASS; 3008 } 3009 } 3010 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 3011 pNew->nOut -= nIn; 3012 } 3013 } 3014 if( nOut==0 ) 3015 #endif 3016 { 3017 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 3018 if( eOp & WO_ISNULL ){ 3019 /* TUNING: If there is no likelihood() value, assume that a 3020 ** "col IS NULL" expression matches twice as many rows 3021 ** as (col=?). */ 3022 pNew->nOut += 10; 3023 } 3024 } 3025 } 3026 } 3027 3028 /* Set rCostIdx to the cost of visiting selected rows in index. Add 3029 ** it to pNew->rRun, which is currently set to the cost of the index 3030 ** seek only. Then, if this is a non-covering index, add the cost of 3031 ** visiting the rows in the main table. */ 3032 assert( pSrc->pTab->szTabRow>0 ); 3033 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 3034 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 3035 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 3036 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 3037 } 3038 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 3039 3040 nOutUnadjusted = pNew->nOut; 3041 pNew->rRun += nInMul + nIn; 3042 pNew->nOut += nInMul + nIn; 3043 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 3044 rc = whereLoopInsert(pBuilder, pNew); 3045 3046 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 3047 pNew->nOut = saved_nOut; 3048 }else{ 3049 pNew->nOut = nOutUnadjusted; 3050 } 3051 3052 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 3053 && pNew->u.btree.nEq<pProbe->nColumn 3054 && (pNew->u.btree.nEq<pProbe->nKeyCol || 3055 pProbe->idxType!=SQLITE_IDXTYPE_PRIMARYKEY) 3056 ){ 3057 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 3058 } 3059 pNew->nOut = saved_nOut; 3060 #ifdef SQLITE_ENABLE_STAT4 3061 pBuilder->nRecValid = nRecValid; 3062 #endif 3063 } 3064 pNew->prereq = saved_prereq; 3065 pNew->u.btree.nEq = saved_nEq; 3066 pNew->u.btree.nBtm = saved_nBtm; 3067 pNew->u.btree.nTop = saved_nTop; 3068 pNew->nSkip = saved_nSkip; 3069 pNew->wsFlags = saved_wsFlags; 3070 pNew->nOut = saved_nOut; 3071 pNew->nLTerm = saved_nLTerm; 3072 3073 /* Consider using a skip-scan if there are no WHERE clause constraints 3074 ** available for the left-most terms of the index, and if the average 3075 ** number of repeats in the left-most terms is at least 18. 3076 ** 3077 ** The magic number 18 is selected on the basis that scanning 17 rows 3078 ** is almost always quicker than an index seek (even though if the index 3079 ** contains fewer than 2^17 rows we assume otherwise in other parts of 3080 ** the code). And, even if it is not, it should not be too much slower. 3081 ** On the other hand, the extra seeks could end up being significantly 3082 ** more expensive. */ 3083 assert( 42==sqlite3LogEst(18) ); 3084 if( saved_nEq==saved_nSkip 3085 && saved_nEq+1<pProbe->nKeyCol 3086 && saved_nEq==pNew->nLTerm 3087 && pProbe->noSkipScan==0 3088 && pProbe->hasStat1!=0 3089 && OptimizationEnabled(db, SQLITE_SkipScan) 3090 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 3091 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 3092 ){ 3093 LogEst nIter; 3094 pNew->u.btree.nEq++; 3095 pNew->nSkip++; 3096 pNew->aLTerm[pNew->nLTerm++] = 0; 3097 pNew->wsFlags |= WHERE_SKIPSCAN; 3098 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 3099 pNew->nOut -= nIter; 3100 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 3101 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 3102 nIter += 5; 3103 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 3104 pNew->nOut = saved_nOut; 3105 pNew->u.btree.nEq = saved_nEq; 3106 pNew->nSkip = saved_nSkip; 3107 pNew->wsFlags = saved_wsFlags; 3108 } 3109 3110 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n", 3111 pProbe->pTable->zName, pProbe->zName, saved_nEq, rc)); 3112 return rc; 3113 } 3114 3115 /* 3116 ** Return True if it is possible that pIndex might be useful in 3117 ** implementing the ORDER BY clause in pBuilder. 3118 ** 3119 ** Return False if pBuilder does not contain an ORDER BY clause or 3120 ** if there is no way for pIndex to be useful in implementing that 3121 ** ORDER BY clause. 3122 */ 3123 static int indexMightHelpWithOrderBy( 3124 WhereLoopBuilder *pBuilder, 3125 Index *pIndex, 3126 int iCursor 3127 ){ 3128 ExprList *pOB; 3129 ExprList *aColExpr; 3130 int ii, jj; 3131 3132 if( pIndex->bUnordered ) return 0; 3133 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 3134 for(ii=0; ii<pOB->nExpr; ii++){ 3135 Expr *pExpr = sqlite3ExprSkipCollateAndLikely(pOB->a[ii].pExpr); 3136 if( NEVER(pExpr==0) ) continue; 3137 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 3138 if( pExpr->iColumn<0 ) return 1; 3139 for(jj=0; jj<pIndex->nKeyCol; jj++){ 3140 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 3141 } 3142 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 3143 for(jj=0; jj<pIndex->nKeyCol; jj++){ 3144 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 3145 if( sqlite3ExprCompareSkip(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 3146 return 1; 3147 } 3148 } 3149 } 3150 } 3151 return 0; 3152 } 3153 3154 /* Check to see if a partial index with pPartIndexWhere can be used 3155 ** in the current query. Return true if it can be and false if not. 3156 */ 3157 static int whereUsablePartialIndex( 3158 int iTab, /* The table for which we want an index */ 3159 int isLeft, /* True if iTab is the right table of a LEFT JOIN */ 3160 WhereClause *pWC, /* The WHERE clause of the query */ 3161 Expr *pWhere /* The WHERE clause from the partial index */ 3162 ){ 3163 int i; 3164 WhereTerm *pTerm; 3165 Parse *pParse = pWC->pWInfo->pParse; 3166 while( pWhere->op==TK_AND ){ 3167 if( !whereUsablePartialIndex(iTab,isLeft,pWC,pWhere->pLeft) ) return 0; 3168 pWhere = pWhere->pRight; 3169 } 3170 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 3171 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 3172 Expr *pExpr; 3173 pExpr = pTerm->pExpr; 3174 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->w.iRightJoinTable==iTab) 3175 && (isLeft==0 || ExprHasProperty(pExpr, EP_FromJoin)) 3176 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 3177 && (pTerm->wtFlags & TERM_VNULL)==0 3178 ){ 3179 return 1; 3180 } 3181 } 3182 return 0; 3183 } 3184 3185 /* 3186 ** Add all WhereLoop objects for a single table of the join where the table 3187 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 3188 ** a b-tree table, not a virtual table. 3189 ** 3190 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 3191 ** are calculated as follows: 3192 ** 3193 ** For a full scan, assuming the table (or index) contains nRow rows: 3194 ** 3195 ** cost = nRow * 3.0 // full-table scan 3196 ** cost = nRow * K // scan of covering index 3197 ** cost = nRow * (K+3.0) // scan of non-covering index 3198 ** 3199 ** where K is a value between 1.1 and 3.0 set based on the relative 3200 ** estimated average size of the index and table records. 3201 ** 3202 ** For an index scan, where nVisit is the number of index rows visited 3203 ** by the scan, and nSeek is the number of seek operations required on 3204 ** the index b-tree: 3205 ** 3206 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 3207 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 3208 ** 3209 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 3210 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 3211 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 3212 ** 3213 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 3214 ** of uncertainty. For this reason, scoring is designed to pick plans that 3215 ** "do the least harm" if the estimates are inaccurate. For example, a 3216 ** log(nRow) factor is omitted from a non-covering index scan in order to 3217 ** bias the scoring in favor of using an index, since the worst-case 3218 ** performance of using an index is far better than the worst-case performance 3219 ** of a full table scan. 3220 */ 3221 static int whereLoopAddBtree( 3222 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3223 Bitmask mPrereq /* Extra prerequesites for using this table */ 3224 ){ 3225 WhereInfo *pWInfo; /* WHERE analysis context */ 3226 Index *pProbe; /* An index we are evaluating */ 3227 Index sPk; /* A fake index object for the primary key */ 3228 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 3229 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 3230 SrcList *pTabList; /* The FROM clause */ 3231 SrcItem *pSrc; /* The FROM clause btree term to add */ 3232 WhereLoop *pNew; /* Template WhereLoop object */ 3233 int rc = SQLITE_OK; /* Return code */ 3234 int iSortIdx = 1; /* Index number */ 3235 int b; /* A boolean value */ 3236 LogEst rSize; /* number of rows in the table */ 3237 WhereClause *pWC; /* The parsed WHERE clause */ 3238 Table *pTab; /* Table being queried */ 3239 3240 pNew = pBuilder->pNew; 3241 pWInfo = pBuilder->pWInfo; 3242 pTabList = pWInfo->pTabList; 3243 pSrc = pTabList->a + pNew->iTab; 3244 pTab = pSrc->pTab; 3245 pWC = pBuilder->pWC; 3246 assert( !IsVirtual(pSrc->pTab) ); 3247 3248 if( pSrc->fg.isIndexedBy ){ 3249 assert( pSrc->fg.isCte==0 ); 3250 /* An INDEXED BY clause specifies a particular index to use */ 3251 pProbe = pSrc->u2.pIBIndex; 3252 }else if( !HasRowid(pTab) ){ 3253 pProbe = pTab->pIndex; 3254 }else{ 3255 /* There is no INDEXED BY clause. Create a fake Index object in local 3256 ** variable sPk to represent the rowid primary key index. Make this 3257 ** fake index the first in a chain of Index objects with all of the real 3258 ** indices to follow */ 3259 Index *pFirst; /* First of real indices on the table */ 3260 memset(&sPk, 0, sizeof(Index)); 3261 sPk.nKeyCol = 1; 3262 sPk.nColumn = 1; 3263 sPk.aiColumn = &aiColumnPk; 3264 sPk.aiRowLogEst = aiRowEstPk; 3265 sPk.onError = OE_Replace; 3266 sPk.pTable = pTab; 3267 sPk.szIdxRow = pTab->szTabRow; 3268 sPk.idxType = SQLITE_IDXTYPE_IPK; 3269 aiRowEstPk[0] = pTab->nRowLogEst; 3270 aiRowEstPk[1] = 0; 3271 pFirst = pSrc->pTab->pIndex; 3272 if( pSrc->fg.notIndexed==0 ){ 3273 /* The real indices of the table are only considered if the 3274 ** NOT INDEXED qualifier is omitted from the FROM clause */ 3275 sPk.pNext = pFirst; 3276 } 3277 pProbe = &sPk; 3278 } 3279 rSize = pTab->nRowLogEst; 3280 3281 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 3282 /* Automatic indexes */ 3283 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 3284 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 3285 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 3286 && !pSrc->fg.isIndexedBy /* Has no INDEXED BY clause */ 3287 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 3288 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 3289 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 3290 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 3291 ){ 3292 /* Generate auto-index WhereLoops */ 3293 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 3294 WhereTerm *pTerm; 3295 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 3296 rLogSize = estLog(rSize); 3297 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 3298 if( pTerm->prereqRight & pNew->maskSelf ) continue; 3299 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 3300 pNew->u.btree.nEq = 1; 3301 pNew->nSkip = 0; 3302 pNew->u.btree.pIndex = 0; 3303 pNew->nLTerm = 1; 3304 pNew->aLTerm[0] = pTerm; 3305 /* TUNING: One-time cost for computing the automatic index is 3306 ** estimated to be X*N*log2(N) where N is the number of rows in 3307 ** the table being indexed and where X is 7 (LogEst=28) for normal 3308 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value 3309 ** of X is smaller for views and subqueries so that the query planner 3310 ** will be more aggressive about generating automatic indexes for 3311 ** those objects, since there is no opportunity to add schema 3312 ** indexes on subqueries and views. */ 3313 pNew->rSetup = rLogSize + rSize; 3314 if( !IsView(pTab) && (pTab->tabFlags & TF_Ephemeral)==0 ){ 3315 pNew->rSetup += 28; 3316 }else{ 3317 pNew->rSetup -= 10; 3318 } 3319 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 3320 if( pNew->rSetup<0 ) pNew->rSetup = 0; 3321 /* TUNING: Each index lookup yields 20 rows in the table. This 3322 ** is more than the usual guess of 10 rows, since we have no way 3323 ** of knowing how selective the index will ultimately be. It would 3324 ** not be unreasonable to make this value much larger. */ 3325 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 3326 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 3327 pNew->wsFlags = WHERE_AUTO_INDEX; 3328 pNew->prereq = mPrereq | pTerm->prereqRight; 3329 rc = whereLoopInsert(pBuilder, pNew); 3330 } 3331 } 3332 } 3333 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 3334 3335 /* Loop over all indices. If there was an INDEXED BY clause, then only 3336 ** consider index pProbe. */ 3337 for(; rc==SQLITE_OK && pProbe; 3338 pProbe=(pSrc->fg.isIndexedBy ? 0 : pProbe->pNext), iSortIdx++ 3339 ){ 3340 int isLeft = (pSrc->fg.jointype & JT_OUTER)!=0; 3341 if( pProbe->pPartIdxWhere!=0 3342 && !whereUsablePartialIndex(pSrc->iCursor, isLeft, pWC, 3343 pProbe->pPartIdxWhere) 3344 ){ 3345 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 3346 continue; /* Partial index inappropriate for this query */ 3347 } 3348 if( pProbe->bNoQuery ) continue; 3349 rSize = pProbe->aiRowLogEst[0]; 3350 pNew->u.btree.nEq = 0; 3351 pNew->u.btree.nBtm = 0; 3352 pNew->u.btree.nTop = 0; 3353 pNew->nSkip = 0; 3354 pNew->nLTerm = 0; 3355 pNew->iSortIdx = 0; 3356 pNew->rSetup = 0; 3357 pNew->prereq = mPrereq; 3358 pNew->nOut = rSize; 3359 pNew->u.btree.pIndex = pProbe; 3360 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 3361 3362 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 3363 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 3364 if( pProbe->idxType==SQLITE_IDXTYPE_IPK ){ 3365 /* Integer primary key index */ 3366 pNew->wsFlags = WHERE_IPK; 3367 3368 /* Full table scan */ 3369 pNew->iSortIdx = b ? iSortIdx : 0; 3370 /* TUNING: Cost of full table scan is 3.0*N. The 3.0 factor is an 3371 ** extra cost designed to discourage the use of full table scans, 3372 ** since index lookups have better worst-case performance if our 3373 ** stat guesses are wrong. Reduce the 3.0 penalty slightly 3374 ** (to 2.75) if we have valid STAT4 information for the table. 3375 ** At 2.75, a full table scan is preferred over using an index on 3376 ** a column with just two distinct values where each value has about 3377 ** an equal number of appearances. Without STAT4 data, we still want 3378 ** to use an index in that case, since the constraint might be for 3379 ** the scarcer of the two values, and in that case an index lookup is 3380 ** better. 3381 */ 3382 #ifdef SQLITE_ENABLE_STAT4 3383 pNew->rRun = rSize + 16 - 2*((pTab->tabFlags & TF_HasStat4)!=0); 3384 #else 3385 pNew->rRun = rSize + 16; 3386 #endif 3387 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3388 whereLoopOutputAdjust(pWC, pNew, rSize); 3389 rc = whereLoopInsert(pBuilder, pNew); 3390 pNew->nOut = rSize; 3391 if( rc ) break; 3392 }else{ 3393 Bitmask m; 3394 if( pProbe->isCovering ){ 3395 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 3396 m = 0; 3397 }else{ 3398 m = pSrc->colUsed & pProbe->colNotIdxed; 3399 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 3400 } 3401 3402 /* Full scan via index */ 3403 if( b 3404 || !HasRowid(pTab) 3405 || pProbe->pPartIdxWhere!=0 3406 || pSrc->fg.isIndexedBy 3407 || ( m==0 3408 && pProbe->bUnordered==0 3409 && (pProbe->szIdxRow<pTab->szTabRow) 3410 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 3411 && sqlite3GlobalConfig.bUseCis 3412 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 3413 ) 3414 ){ 3415 pNew->iSortIdx = b ? iSortIdx : 0; 3416 3417 /* The cost of visiting the index rows is N*K, where K is 3418 ** between 1.1 and 3.0, depending on the relative sizes of the 3419 ** index and table rows. */ 3420 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 3421 if( m!=0 ){ 3422 /* If this is a non-covering index scan, add in the cost of 3423 ** doing table lookups. The cost will be 3x the number of 3424 ** lookups. Take into account WHERE clause terms that can be 3425 ** satisfied using just the index, and that do not require a 3426 ** table lookup. */ 3427 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 3428 int ii; 3429 int iCur = pSrc->iCursor; 3430 WhereClause *pWC2 = &pWInfo->sWC; 3431 for(ii=0; ii<pWC2->nTerm; ii++){ 3432 WhereTerm *pTerm = &pWC2->a[ii]; 3433 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 3434 break; 3435 } 3436 /* pTerm can be evaluated using just the index. So reduce 3437 ** the expected number of table lookups accordingly */ 3438 if( pTerm->truthProb<=0 ){ 3439 nLookup += pTerm->truthProb; 3440 }else{ 3441 nLookup--; 3442 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 3443 } 3444 } 3445 3446 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 3447 } 3448 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3449 whereLoopOutputAdjust(pWC, pNew, rSize); 3450 rc = whereLoopInsert(pBuilder, pNew); 3451 pNew->nOut = rSize; 3452 if( rc ) break; 3453 } 3454 } 3455 3456 pBuilder->bldFlags1 = 0; 3457 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 3458 if( pBuilder->bldFlags1==SQLITE_BLDF1_INDEXED ){ 3459 /* If a non-unique index is used, or if a prefix of the key for 3460 ** unique index is used (making the index functionally non-unique) 3461 ** then the sqlite_stat1 data becomes important for scoring the 3462 ** plan */ 3463 pTab->tabFlags |= TF_StatsUsed; 3464 } 3465 #ifdef SQLITE_ENABLE_STAT4 3466 sqlite3Stat4ProbeFree(pBuilder->pRec); 3467 pBuilder->nRecValid = 0; 3468 pBuilder->pRec = 0; 3469 #endif 3470 } 3471 return rc; 3472 } 3473 3474 #ifndef SQLITE_OMIT_VIRTUALTABLE 3475 3476 /* 3477 ** Return true if pTerm is a virtual table LIMIT or OFFSET term. 3478 */ 3479 static int isLimitTerm(WhereTerm *pTerm){ 3480 assert( pTerm->eOperator==WO_AUX || pTerm->eMatchOp==0 ); 3481 return pTerm->eMatchOp>=SQLITE_INDEX_CONSTRAINT_LIMIT 3482 && pTerm->eMatchOp<=SQLITE_INDEX_CONSTRAINT_OFFSET; 3483 } 3484 3485 /* 3486 ** Argument pIdxInfo is already populated with all constraints that may 3487 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 3488 ** function marks a subset of those constraints usable, invokes the 3489 ** xBestIndex method and adds the returned plan to pBuilder. 3490 ** 3491 ** A constraint is marked usable if: 3492 ** 3493 ** * Argument mUsable indicates that its prerequisites are available, and 3494 ** 3495 ** * It is not one of the operators specified in the mExclude mask passed 3496 ** as the fourth argument (which in practice is either WO_IN or 0). 3497 ** 3498 ** Argument mPrereq is a mask of tables that must be scanned before the 3499 ** virtual table in question. These are added to the plans prerequisites 3500 ** before it is added to pBuilder. 3501 ** 3502 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3503 ** uses one or more WO_IN terms, or false otherwise. 3504 */ 3505 static int whereLoopAddVirtualOne( 3506 WhereLoopBuilder *pBuilder, 3507 Bitmask mPrereq, /* Mask of tables that must be used. */ 3508 Bitmask mUsable, /* Mask of usable tables */ 3509 u16 mExclude, /* Exclude terms using these operators */ 3510 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3511 u16 mNoOmit, /* Do not omit these constraints */ 3512 int *pbIn, /* OUT: True if plan uses an IN(...) op */ 3513 int *pbRetryLimit /* OUT: Retry without LIMIT/OFFSET */ 3514 ){ 3515 WhereClause *pWC = pBuilder->pWC; 3516 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3517 struct sqlite3_index_constraint *pIdxCons; 3518 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3519 int i; 3520 int mxTerm; 3521 int rc = SQLITE_OK; 3522 WhereLoop *pNew = pBuilder->pNew; 3523 Parse *pParse = pBuilder->pWInfo->pParse; 3524 SrcItem *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3525 int nConstraint = pIdxInfo->nConstraint; 3526 3527 assert( (mUsable & mPrereq)==mPrereq ); 3528 *pbIn = 0; 3529 pNew->prereq = mPrereq; 3530 3531 /* Set the usable flag on the subset of constraints identified by 3532 ** arguments mUsable and mExclude. */ 3533 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3534 for(i=0; i<nConstraint; i++, pIdxCons++){ 3535 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3536 pIdxCons->usable = 0; 3537 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3538 && (pTerm->eOperator & mExclude)==0 3539 && (pbRetryLimit || !isLimitTerm(pTerm)) 3540 ){ 3541 pIdxCons->usable = 1; 3542 } 3543 } 3544 3545 /* Initialize the output fields of the sqlite3_index_info structure */ 3546 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3547 assert( pIdxInfo->needToFreeIdxStr==0 ); 3548 pIdxInfo->idxStr = 0; 3549 pIdxInfo->idxNum = 0; 3550 pIdxInfo->orderByConsumed = 0; 3551 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3552 pIdxInfo->estimatedRows = 25; 3553 pIdxInfo->idxFlags = 0; 3554 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3555 pHidden->mHandleIn = 0; 3556 3557 /* Invoke the virtual table xBestIndex() method */ 3558 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3559 if( rc ){ 3560 if( rc==SQLITE_CONSTRAINT ){ 3561 /* If the xBestIndex method returns SQLITE_CONSTRAINT, that means 3562 ** that the particular combination of parameters provided is unusable. 3563 ** Make no entries in the loop table. 3564 */ 3565 WHERETRACE(0xffff, (" ^^^^--- non-viable plan rejected!\n")); 3566 return SQLITE_OK; 3567 } 3568 return rc; 3569 } 3570 3571 mxTerm = -1; 3572 assert( pNew->nLSlot>=nConstraint ); 3573 memset(pNew->aLTerm, 0, sizeof(pNew->aLTerm[0])*nConstraint ); 3574 memset(&pNew->u.vtab, 0, sizeof(pNew->u.vtab)); 3575 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3576 for(i=0; i<nConstraint; i++, pIdxCons++){ 3577 int iTerm; 3578 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3579 WhereTerm *pTerm; 3580 int j = pIdxCons->iTermOffset; 3581 if( iTerm>=nConstraint 3582 || j<0 3583 || j>=pWC->nTerm 3584 || pNew->aLTerm[iTerm]!=0 3585 || pIdxCons->usable==0 3586 ){ 3587 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3588 testcase( pIdxInfo->needToFreeIdxStr ); 3589 return SQLITE_ERROR; 3590 } 3591 testcase( iTerm==nConstraint-1 ); 3592 testcase( j==0 ); 3593 testcase( j==pWC->nTerm-1 ); 3594 pTerm = &pWC->a[j]; 3595 pNew->prereq |= pTerm->prereqRight; 3596 assert( iTerm<pNew->nLSlot ); 3597 pNew->aLTerm[iTerm] = pTerm; 3598 if( iTerm>mxTerm ) mxTerm = iTerm; 3599 testcase( iTerm==15 ); 3600 testcase( iTerm==16 ); 3601 if( pUsage[i].omit ){ 3602 if( i<16 && ((1<<i)&mNoOmit)==0 ){ 3603 testcase( i!=iTerm ); 3604 pNew->u.vtab.omitMask |= 1<<iTerm; 3605 }else{ 3606 testcase( i!=iTerm ); 3607 } 3608 if( pTerm->eMatchOp==SQLITE_INDEX_CONSTRAINT_OFFSET ){ 3609 pNew->u.vtab.bOmitOffset = 1; 3610 } 3611 } 3612 if( SMASKBIT32(i) & pHidden->mHandleIn ){ 3613 pNew->u.vtab.mHandleIn |= MASKBIT32(iTerm); 3614 }else if( (pTerm->eOperator & WO_IN)!=0 ){ 3615 /* A virtual table that is constrained by an IN clause may not 3616 ** consume the ORDER BY clause because (1) the order of IN terms 3617 ** is not necessarily related to the order of output terms and 3618 ** (2) Multiple outputs from a single IN value will not merge 3619 ** together. */ 3620 pIdxInfo->orderByConsumed = 0; 3621 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3622 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3623 } 3624 3625 if( isLimitTerm(pTerm) && *pbIn ){ 3626 /* If there is an IN(...) term handled as an == (separate call to 3627 ** xFilter for each value on the RHS of the IN) and a LIMIT or 3628 ** OFFSET term handled as well, the plan is unusable. Set output 3629 ** variable *pbRetryLimit to true to tell the caller to retry with 3630 ** LIMIT and OFFSET disabled. */ 3631 if( pIdxInfo->needToFreeIdxStr ){ 3632 sqlite3_free(pIdxInfo->idxStr); 3633 pIdxInfo->idxStr = 0; 3634 pIdxInfo->needToFreeIdxStr = 0; 3635 } 3636 *pbRetryLimit = 1; 3637 return SQLITE_OK; 3638 } 3639 } 3640 } 3641 3642 pNew->nLTerm = mxTerm+1; 3643 for(i=0; i<=mxTerm; i++){ 3644 if( pNew->aLTerm[i]==0 ){ 3645 /* The non-zero argvIdx values must be contiguous. Raise an 3646 ** error if they are not */ 3647 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3648 testcase( pIdxInfo->needToFreeIdxStr ); 3649 return SQLITE_ERROR; 3650 } 3651 } 3652 assert( pNew->nLTerm<=pNew->nLSlot ); 3653 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3654 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3655 pIdxInfo->needToFreeIdxStr = 0; 3656 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3657 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3658 pIdxInfo->nOrderBy : 0); 3659 pNew->rSetup = 0; 3660 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3661 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3662 3663 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3664 ** that the scan will visit at most one row. Clear it otherwise. */ 3665 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3666 pNew->wsFlags |= WHERE_ONEROW; 3667 }else{ 3668 pNew->wsFlags &= ~WHERE_ONEROW; 3669 } 3670 rc = whereLoopInsert(pBuilder, pNew); 3671 if( pNew->u.vtab.needFree ){ 3672 sqlite3_free(pNew->u.vtab.idxStr); 3673 pNew->u.vtab.needFree = 0; 3674 } 3675 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3676 *pbIn, (sqlite3_uint64)mPrereq, 3677 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3678 3679 return rc; 3680 } 3681 3682 /* 3683 ** Return the collating sequence for a constraint passed into xBestIndex. 3684 ** 3685 ** pIdxInfo must be an sqlite3_index_info structure passed into xBestIndex. 3686 ** This routine depends on there being a HiddenIndexInfo structure immediately 3687 ** following the sqlite3_index_info structure. 3688 ** 3689 ** Return a pointer to the collation name: 3690 ** 3691 ** 1. If there is an explicit COLLATE operator on the constaint, return it. 3692 ** 3693 ** 2. Else, if the column has an alternative collation, return that. 3694 ** 3695 ** 3. Otherwise, return "BINARY". 3696 */ 3697 const char *sqlite3_vtab_collation(sqlite3_index_info *pIdxInfo, int iCons){ 3698 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3699 const char *zRet = 0; 3700 if( iCons>=0 && iCons<pIdxInfo->nConstraint ){ 3701 CollSeq *pC = 0; 3702 int iTerm = pIdxInfo->aConstraint[iCons].iTermOffset; 3703 Expr *pX = pHidden->pWC->a[iTerm].pExpr; 3704 if( pX->pLeft ){ 3705 pC = sqlite3ExprCompareCollSeq(pHidden->pParse, pX); 3706 } 3707 zRet = (pC ? pC->zName : sqlite3StrBINARY); 3708 } 3709 return zRet; 3710 } 3711 3712 /* 3713 ** Return true if constraint iCons is really an IN(...) constraint, or 3714 ** false otherwise. If iCons is an IN(...) constraint, set (if bHandle!=0) 3715 ** or clear (if bHandle==0) the flag to handle it using an iterator. 3716 */ 3717 int sqlite3_vtab_in(sqlite3_index_info *pIdxInfo, int iCons, int bHandle){ 3718 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3719 u32 m = SMASKBIT32(iCons); 3720 if( m & pHidden->mIn ){ 3721 if( bHandle==0 ){ 3722 pHidden->mHandleIn &= ~m; 3723 }else if( bHandle>0 ){ 3724 pHidden->mHandleIn |= m; 3725 } 3726 return 1; 3727 } 3728 return 0; 3729 } 3730 3731 /* 3732 ** This interface is callable from within the xBestIndex callback only. 3733 ** 3734 ** If possible, set (*ppVal) to point to an object containing the value 3735 ** on the right-hand-side of constraint iCons. 3736 */ 3737 int sqlite3_vtab_rhs_value( 3738 sqlite3_index_info *pIdxInfo, /* Copy of first argument to xBestIndex */ 3739 int iCons, /* Constraint for which RHS is wanted */ 3740 sqlite3_value **ppVal /* Write value extracted here */ 3741 ){ 3742 HiddenIndexInfo *pH = (HiddenIndexInfo*)&pIdxInfo[1]; 3743 sqlite3_value *pVal = 0; 3744 int rc = SQLITE_OK; 3745 if( iCons<0 || iCons>=pIdxInfo->nConstraint ){ 3746 rc = SQLITE_MISUSE; /* EV: R-30545-25046 */ 3747 }else{ 3748 if( pH->aRhs[iCons]==0 ){ 3749 WhereTerm *pTerm = &pH->pWC->a[pIdxInfo->aConstraint[iCons].iTermOffset]; 3750 rc = sqlite3ValueFromExpr( 3751 pH->pParse->db, pTerm->pExpr->pRight, ENC(pH->pParse->db), 3752 SQLITE_AFF_BLOB, &pH->aRhs[iCons] 3753 ); 3754 testcase( rc!=SQLITE_OK ); 3755 } 3756 pVal = pH->aRhs[iCons]; 3757 } 3758 *ppVal = pVal; 3759 3760 if( rc==SQLITE_OK && pVal==0 ){ /* IMP: R-19933-32160 */ 3761 rc = SQLITE_NOTFOUND; /* IMP: R-36424-56542 */ 3762 } 3763 3764 return rc; 3765 } 3766 3767 /* 3768 ** Return true if ORDER BY clause may be handled as DISTINCT. 3769 */ 3770 int sqlite3_vtab_distinct(sqlite3_index_info *pIdxInfo){ 3771 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3772 assert( pHidden->eDistinct==0 3773 || pHidden->eDistinct==1 3774 || pHidden->eDistinct==2 ); 3775 return pHidden->eDistinct; 3776 } 3777 3778 #if (defined(SQLITE_ENABLE_DBPAGE_VTAB) || defined(SQLITE_TEST)) \ 3779 && !defined(SQLITE_OMIT_VIRTUALTABLE) 3780 /* 3781 ** Cause the prepared statement that is associated with a call to 3782 ** xBestIndex to open write transactions on all attached schemas. 3783 ** This is used by the (built-in) sqlite_dbpage virtual table. 3784 */ 3785 void sqlite3VtabWriteAll(sqlite3_index_info *pIdxInfo){ 3786 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3787 Parse *pParse = pHidden->pParse; 3788 int nDb = pParse->db->nDb; 3789 int i; 3790 for(i=0; i<nDb; i++) sqlite3BeginWriteOperation(pParse, 0, i); 3791 } 3792 #endif 3793 3794 /* 3795 ** Add all WhereLoop objects for a table of the join identified by 3796 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3797 ** 3798 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3799 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3800 ** entries that occur before the virtual table in the FROM clause and are 3801 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3802 ** mUnusable mask contains all FROM clause entries that occur after the 3803 ** virtual table and are separated from it by at least one LEFT or 3804 ** CROSS JOIN. 3805 ** 3806 ** For example, if the query were: 3807 ** 3808 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3809 ** 3810 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3811 ** 3812 ** All the tables in mPrereq must be scanned before the current virtual 3813 ** table. So any terms for which all prerequisites are satisfied by 3814 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3815 ** Conversely, all tables in mUnusable must be scanned after the current 3816 ** virtual table, so any terms for which the prerequisites overlap with 3817 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3818 */ 3819 static int whereLoopAddVirtual( 3820 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3821 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3822 Bitmask mUnusable /* Tables that must be scanned after this one */ 3823 ){ 3824 int rc = SQLITE_OK; /* Return code */ 3825 WhereInfo *pWInfo; /* WHERE analysis context */ 3826 Parse *pParse; /* The parsing context */ 3827 WhereClause *pWC; /* The WHERE clause */ 3828 SrcItem *pSrc; /* The FROM clause term to search */ 3829 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3830 int nConstraint; /* Number of constraints in p */ 3831 int bIn; /* True if plan uses IN(...) operator */ 3832 WhereLoop *pNew; 3833 Bitmask mBest; /* Tables used by best possible plan */ 3834 u16 mNoOmit; 3835 int bRetry = 0; /* True to retry with LIMIT/OFFSET disabled */ 3836 3837 assert( (mPrereq & mUnusable)==0 ); 3838 pWInfo = pBuilder->pWInfo; 3839 pParse = pWInfo->pParse; 3840 pWC = pBuilder->pWC; 3841 pNew = pBuilder->pNew; 3842 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3843 assert( IsVirtual(pSrc->pTab) ); 3844 p = allocateIndexInfo(pWInfo, pWC, mUnusable, pSrc, &mNoOmit); 3845 if( p==0 ) return SQLITE_NOMEM_BKPT; 3846 pNew->rSetup = 0; 3847 pNew->wsFlags = WHERE_VIRTUALTABLE; 3848 pNew->nLTerm = 0; 3849 pNew->u.vtab.needFree = 0; 3850 nConstraint = p->nConstraint; 3851 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3852 freeIndexInfo(pParse->db, p); 3853 return SQLITE_NOMEM_BKPT; 3854 } 3855 3856 /* First call xBestIndex() with all constraints usable. */ 3857 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc->pTab->zName)); 3858 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3859 rc = whereLoopAddVirtualOne( 3860 pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn, &bRetry 3861 ); 3862 if( bRetry ){ 3863 assert( rc==SQLITE_OK ); 3864 rc = whereLoopAddVirtualOne( 3865 pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn, 0 3866 ); 3867 } 3868 3869 /* If the call to xBestIndex() with all terms enabled produced a plan 3870 ** that does not require any source tables (IOW: a plan with mBest==0) 3871 ** and does not use an IN(...) operator, then there is no point in making 3872 ** any further calls to xBestIndex() since they will all return the same 3873 ** result (if the xBestIndex() implementation is sane). */ 3874 if( rc==SQLITE_OK && ((mBest = (pNew->prereq & ~mPrereq))!=0 || bIn) ){ 3875 int seenZero = 0; /* True if a plan with no prereqs seen */ 3876 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3877 Bitmask mPrev = 0; 3878 Bitmask mBestNoIn = 0; 3879 3880 /* If the plan produced by the earlier call uses an IN(...) term, call 3881 ** xBestIndex again, this time with IN(...) terms disabled. */ 3882 if( bIn ){ 3883 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3884 rc = whereLoopAddVirtualOne( 3885 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn, 0); 3886 assert( bIn==0 ); 3887 mBestNoIn = pNew->prereq & ~mPrereq; 3888 if( mBestNoIn==0 ){ 3889 seenZero = 1; 3890 seenZeroNoIN = 1; 3891 } 3892 } 3893 3894 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3895 ** in the set of terms that apply to the current virtual table. */ 3896 while( rc==SQLITE_OK ){ 3897 int i; 3898 Bitmask mNext = ALLBITS; 3899 assert( mNext>0 ); 3900 for(i=0; i<nConstraint; i++){ 3901 Bitmask mThis = ( 3902 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3903 ); 3904 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3905 } 3906 mPrev = mNext; 3907 if( mNext==ALLBITS ) break; 3908 if( mNext==mBest || mNext==mBestNoIn ) continue; 3909 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3910 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3911 rc = whereLoopAddVirtualOne( 3912 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn, 0); 3913 if( pNew->prereq==mPrereq ){ 3914 seenZero = 1; 3915 if( bIn==0 ) seenZeroNoIN = 1; 3916 } 3917 } 3918 3919 /* If the calls to xBestIndex() in the above loop did not find a plan 3920 ** that requires no source tables at all (i.e. one guaranteed to be 3921 ** usable), make a call here with all source tables disabled */ 3922 if( rc==SQLITE_OK && seenZero==0 ){ 3923 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3924 rc = whereLoopAddVirtualOne( 3925 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn, 0); 3926 if( bIn==0 ) seenZeroNoIN = 1; 3927 } 3928 3929 /* If the calls to xBestIndex() have so far failed to find a plan 3930 ** that requires no source tables at all and does not use an IN(...) 3931 ** operator, make a final call to obtain one here. */ 3932 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3933 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3934 rc = whereLoopAddVirtualOne( 3935 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn, 0); 3936 } 3937 } 3938 3939 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3940 freeIndexInfo(pParse->db, p); 3941 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc->pTab->zName, rc)); 3942 return rc; 3943 } 3944 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3945 3946 /* 3947 ** Add WhereLoop entries to handle OR terms. This works for either 3948 ** btrees or virtual tables. 3949 */ 3950 static int whereLoopAddOr( 3951 WhereLoopBuilder *pBuilder, 3952 Bitmask mPrereq, 3953 Bitmask mUnusable 3954 ){ 3955 WhereInfo *pWInfo = pBuilder->pWInfo; 3956 WhereClause *pWC; 3957 WhereLoop *pNew; 3958 WhereTerm *pTerm, *pWCEnd; 3959 int rc = SQLITE_OK; 3960 int iCur; 3961 WhereClause tempWC; 3962 WhereLoopBuilder sSubBuild; 3963 WhereOrSet sSum, sCur; 3964 SrcItem *pItem; 3965 3966 pWC = pBuilder->pWC; 3967 pWCEnd = pWC->a + pWC->nTerm; 3968 pNew = pBuilder->pNew; 3969 memset(&sSum, 0, sizeof(sSum)); 3970 pItem = pWInfo->pTabList->a + pNew->iTab; 3971 iCur = pItem->iCursor; 3972 3973 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3974 if( (pTerm->eOperator & WO_OR)!=0 3975 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3976 ){ 3977 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3978 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3979 WhereTerm *pOrTerm; 3980 int once = 1; 3981 int i, j; 3982 3983 sSubBuild = *pBuilder; 3984 sSubBuild.pOrSet = &sCur; 3985 3986 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3987 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3988 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3989 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3990 }else if( pOrTerm->leftCursor==iCur ){ 3991 tempWC.pWInfo = pWC->pWInfo; 3992 tempWC.pOuter = pWC; 3993 tempWC.op = TK_AND; 3994 tempWC.nTerm = 1; 3995 tempWC.nBase = 1; 3996 tempWC.a = pOrTerm; 3997 sSubBuild.pWC = &tempWC; 3998 }else{ 3999 continue; 4000 } 4001 sCur.n = 0; 4002 #ifdef WHERETRACE_ENABLED 4003 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 4004 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 4005 if( sqlite3WhereTrace & 0x400 ){ 4006 sqlite3WhereClausePrint(sSubBuild.pWC); 4007 } 4008 #endif 4009 #ifndef SQLITE_OMIT_VIRTUALTABLE 4010 if( IsVirtual(pItem->pTab) ){ 4011 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 4012 }else 4013 #endif 4014 { 4015 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 4016 } 4017 if( rc==SQLITE_OK ){ 4018 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 4019 } 4020 assert( rc==SQLITE_OK || rc==SQLITE_DONE || sCur.n==0 4021 || rc==SQLITE_NOMEM ); 4022 testcase( rc==SQLITE_NOMEM && sCur.n>0 ); 4023 testcase( rc==SQLITE_DONE ); 4024 if( sCur.n==0 ){ 4025 sSum.n = 0; 4026 break; 4027 }else if( once ){ 4028 whereOrMove(&sSum, &sCur); 4029 once = 0; 4030 }else{ 4031 WhereOrSet sPrev; 4032 whereOrMove(&sPrev, &sSum); 4033 sSum.n = 0; 4034 for(i=0; i<sPrev.n; i++){ 4035 for(j=0; j<sCur.n; j++){ 4036 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 4037 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 4038 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 4039 } 4040 } 4041 } 4042 } 4043 pNew->nLTerm = 1; 4044 pNew->aLTerm[0] = pTerm; 4045 pNew->wsFlags = WHERE_MULTI_OR; 4046 pNew->rSetup = 0; 4047 pNew->iSortIdx = 0; 4048 memset(&pNew->u, 0, sizeof(pNew->u)); 4049 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 4050 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 4051 ** of all sub-scans required by the OR-scan. However, due to rounding 4052 ** errors, it may be that the cost of the OR-scan is equal to its 4053 ** most expensive sub-scan. Add the smallest possible penalty 4054 ** (equivalent to multiplying the cost by 1.07) to ensure that 4055 ** this does not happen. Otherwise, for WHERE clauses such as the 4056 ** following where there is an index on "y": 4057 ** 4058 ** WHERE likelihood(x=?, 0.99) OR y=? 4059 ** 4060 ** the planner may elect to "OR" together a full-table scan and an 4061 ** index lookup. And other similarly odd results. */ 4062 pNew->rRun = sSum.a[i].rRun + 1; 4063 pNew->nOut = sSum.a[i].nOut; 4064 pNew->prereq = sSum.a[i].prereq; 4065 rc = whereLoopInsert(pBuilder, pNew); 4066 } 4067 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 4068 } 4069 } 4070 return rc; 4071 } 4072 4073 /* 4074 ** Add all WhereLoop objects for all tables 4075 */ 4076 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 4077 WhereInfo *pWInfo = pBuilder->pWInfo; 4078 Bitmask mPrereq = 0; 4079 Bitmask mPrior = 0; 4080 int iTab; 4081 SrcList *pTabList = pWInfo->pTabList; 4082 SrcItem *pItem; 4083 SrcItem *pEnd = &pTabList->a[pWInfo->nLevel]; 4084 sqlite3 *db = pWInfo->pParse->db; 4085 int rc = SQLITE_OK; 4086 WhereLoop *pNew; 4087 4088 /* Loop over the tables in the join, from left to right */ 4089 pNew = pBuilder->pNew; 4090 whereLoopInit(pNew); 4091 pBuilder->iPlanLimit = SQLITE_QUERY_PLANNER_LIMIT; 4092 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 4093 Bitmask mUnusable = 0; 4094 pNew->iTab = iTab; 4095 pBuilder->iPlanLimit += SQLITE_QUERY_PLANNER_LIMIT_INCR; 4096 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 4097 if( (pItem->fg.jointype & (JT_LEFT|JT_CROSS))!=0 ){ 4098 /* This condition is true when pItem is the FROM clause term on the 4099 ** right-hand-side of a LEFT or CROSS JOIN. */ 4100 mPrereq = mPrior; 4101 }else{ 4102 mPrereq = 0; 4103 } 4104 #ifndef SQLITE_OMIT_VIRTUALTABLE 4105 if( IsVirtual(pItem->pTab) ){ 4106 SrcItem *p; 4107 for(p=&pItem[1]; p<pEnd; p++){ 4108 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 4109 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 4110 } 4111 } 4112 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 4113 }else 4114 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 4115 { 4116 rc = whereLoopAddBtree(pBuilder, mPrereq); 4117 } 4118 if( rc==SQLITE_OK && pBuilder->pWC->hasOr ){ 4119 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 4120 } 4121 mPrior |= pNew->maskSelf; 4122 if( rc || db->mallocFailed ){ 4123 if( rc==SQLITE_DONE ){ 4124 /* We hit the query planner search limit set by iPlanLimit */ 4125 sqlite3_log(SQLITE_WARNING, "abbreviated query algorithm search"); 4126 rc = SQLITE_OK; 4127 }else{ 4128 break; 4129 } 4130 } 4131 } 4132 4133 whereLoopClear(db, pNew); 4134 return rc; 4135 } 4136 4137 /* 4138 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 4139 ** parameters) to see if it outputs rows in the requested ORDER BY 4140 ** (or GROUP BY) without requiring a separate sort operation. Return N: 4141 ** 4142 ** N>0: N terms of the ORDER BY clause are satisfied 4143 ** N==0: No terms of the ORDER BY clause are satisfied 4144 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 4145 ** 4146 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 4147 ** strict. With GROUP BY and DISTINCT the only requirement is that 4148 ** equivalent rows appear immediately adjacent to one another. GROUP BY 4149 ** and DISTINCT do not require rows to appear in any particular order as long 4150 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 4151 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 4152 ** pOrderBy terms must be matched in strict left-to-right order. 4153 */ 4154 static i8 wherePathSatisfiesOrderBy( 4155 WhereInfo *pWInfo, /* The WHERE clause */ 4156 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 4157 WherePath *pPath, /* The WherePath to check */ 4158 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 4159 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 4160 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 4161 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 4162 ){ 4163 u8 revSet; /* True if rev is known */ 4164 u8 rev; /* Composite sort order */ 4165 u8 revIdx; /* Index sort order */ 4166 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 4167 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 4168 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 4169 u16 eqOpMask; /* Allowed equality operators */ 4170 u16 nKeyCol; /* Number of key columns in pIndex */ 4171 u16 nColumn; /* Total number of ordered columns in the index */ 4172 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 4173 int iLoop; /* Index of WhereLoop in pPath being processed */ 4174 int i, j; /* Loop counters */ 4175 int iCur; /* Cursor number for current WhereLoop */ 4176 int iColumn; /* A column number within table iCur */ 4177 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 4178 WhereTerm *pTerm; /* A single term of the WHERE clause */ 4179 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 4180 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 4181 Index *pIndex; /* The index associated with pLoop */ 4182 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 4183 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 4184 Bitmask obDone; /* Mask of all ORDER BY terms */ 4185 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 4186 Bitmask ready; /* Mask of inner loops */ 4187 4188 /* 4189 ** We say the WhereLoop is "one-row" if it generates no more than one 4190 ** row of output. A WhereLoop is one-row if all of the following are true: 4191 ** (a) All index columns match with WHERE_COLUMN_EQ. 4192 ** (b) The index is unique 4193 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 4194 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 4195 ** 4196 ** We say the WhereLoop is "order-distinct" if the set of columns from 4197 ** that WhereLoop that are in the ORDER BY clause are different for every 4198 ** row of the WhereLoop. Every one-row WhereLoop is automatically 4199 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 4200 ** is not order-distinct. To be order-distinct is not quite the same as being 4201 ** UNIQUE since a UNIQUE column or index can have multiple rows that 4202 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 4203 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 4204 ** 4205 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 4206 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 4207 ** automatically order-distinct. 4208 */ 4209 4210 assert( pOrderBy!=0 ); 4211 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 4212 4213 nOrderBy = pOrderBy->nExpr; 4214 testcase( nOrderBy==BMS-1 ); 4215 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 4216 isOrderDistinct = 1; 4217 obDone = MASKBIT(nOrderBy)-1; 4218 orderDistinctMask = 0; 4219 ready = 0; 4220 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 4221 if( wctrlFlags & (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MAX|WHERE_ORDERBY_MIN) ){ 4222 eqOpMask |= WO_IN; 4223 } 4224 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 4225 if( iLoop>0 ) ready |= pLoop->maskSelf; 4226 if( iLoop<nLoop ){ 4227 pLoop = pPath->aLoop[iLoop]; 4228 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 4229 }else{ 4230 pLoop = pLast; 4231 } 4232 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 4233 if( pLoop->u.vtab.isOrdered && (wctrlFlags & WHERE_DISTINCTBY)==0 ){ 4234 obSat = obDone; 4235 } 4236 break; 4237 }else if( wctrlFlags & WHERE_DISTINCTBY ){ 4238 pLoop->u.btree.nDistinctCol = 0; 4239 } 4240 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 4241 4242 /* Mark off any ORDER BY term X that is a column in the table of 4243 ** the current loop for which there is term in the WHERE 4244 ** clause of the form X IS NULL or X=? that reference only outer 4245 ** loops. 4246 */ 4247 for(i=0; i<nOrderBy; i++){ 4248 if( MASKBIT(i) & obSat ) continue; 4249 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr); 4250 if( NEVER(pOBExpr==0) ) continue; 4251 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue; 4252 if( pOBExpr->iTable!=iCur ) continue; 4253 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 4254 ~ready, eqOpMask, 0); 4255 if( pTerm==0 ) continue; 4256 if( pTerm->eOperator==WO_IN ){ 4257 /* IN terms are only valid for sorting in the ORDER BY LIMIT 4258 ** optimization, and then only if they are actually used 4259 ** by the query plan */ 4260 assert( wctrlFlags & 4261 (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX) ); 4262 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 4263 if( j>=pLoop->nLTerm ) continue; 4264 } 4265 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 4266 Parse *pParse = pWInfo->pParse; 4267 CollSeq *pColl1 = sqlite3ExprNNCollSeq(pParse, pOrderBy->a[i].pExpr); 4268 CollSeq *pColl2 = sqlite3ExprCompareCollSeq(pParse, pTerm->pExpr); 4269 assert( pColl1 ); 4270 if( pColl2==0 || sqlite3StrICmp(pColl1->zName, pColl2->zName) ){ 4271 continue; 4272 } 4273 testcase( pTerm->pExpr->op==TK_IS ); 4274 } 4275 obSat |= MASKBIT(i); 4276 } 4277 4278 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 4279 if( pLoop->wsFlags & WHERE_IPK ){ 4280 pIndex = 0; 4281 nKeyCol = 0; 4282 nColumn = 1; 4283 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 4284 return 0; 4285 }else{ 4286 nKeyCol = pIndex->nKeyCol; 4287 nColumn = pIndex->nColumn; 4288 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 4289 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 4290 || !HasRowid(pIndex->pTable)); 4291 /* All relevant terms of the index must also be non-NULL in order 4292 ** for isOrderDistinct to be true. So the isOrderDistint value 4293 ** computed here might be a false positive. Corrections will be 4294 ** made at tag-20210426-1 below */ 4295 isOrderDistinct = IsUniqueIndex(pIndex) 4296 && (pLoop->wsFlags & WHERE_SKIPSCAN)==0; 4297 } 4298 4299 /* Loop through all columns of the index and deal with the ones 4300 ** that are not constrained by == or IN. 4301 */ 4302 rev = revSet = 0; 4303 distinctColumns = 0; 4304 for(j=0; j<nColumn; j++){ 4305 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 4306 4307 assert( j>=pLoop->u.btree.nEq 4308 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 4309 ); 4310 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 4311 u16 eOp = pLoop->aLTerm[j]->eOperator; 4312 4313 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 4314 ** doing WHERE_ORDERBY_LIMIT processing). Except, IS and ISNULL 4315 ** terms imply that the index is not UNIQUE NOT NULL in which case 4316 ** the loop need to be marked as not order-distinct because it can 4317 ** have repeated NULL rows. 4318 ** 4319 ** If the current term is a column of an ((?,?) IN (SELECT...)) 4320 ** expression for which the SELECT returns more than one column, 4321 ** check that it is the only column used by this loop. Otherwise, 4322 ** if it is one of two or more, none of the columns can be 4323 ** considered to match an ORDER BY term. 4324 */ 4325 if( (eOp & eqOpMask)!=0 ){ 4326 if( eOp & (WO_ISNULL|WO_IS) ){ 4327 testcase( eOp & WO_ISNULL ); 4328 testcase( eOp & WO_IS ); 4329 testcase( isOrderDistinct ); 4330 isOrderDistinct = 0; 4331 } 4332 continue; 4333 }else if( ALWAYS(eOp & WO_IN) ){ 4334 /* ALWAYS() justification: eOp is an equality operator due to the 4335 ** j<pLoop->u.btree.nEq constraint above. Any equality other 4336 ** than WO_IN is captured by the previous "if". So this one 4337 ** always has to be WO_IN. */ 4338 Expr *pX = pLoop->aLTerm[j]->pExpr; 4339 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 4340 if( pLoop->aLTerm[i]->pExpr==pX ){ 4341 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 4342 bOnce = 0; 4343 break; 4344 } 4345 } 4346 } 4347 } 4348 4349 /* Get the column number in the table (iColumn) and sort order 4350 ** (revIdx) for the j-th column of the index. 4351 */ 4352 if( pIndex ){ 4353 iColumn = pIndex->aiColumn[j]; 4354 revIdx = pIndex->aSortOrder[j] & KEYINFO_ORDER_DESC; 4355 if( iColumn==pIndex->pTable->iPKey ) iColumn = XN_ROWID; 4356 }else{ 4357 iColumn = XN_ROWID; 4358 revIdx = 0; 4359 } 4360 4361 /* An unconstrained column that might be NULL means that this 4362 ** WhereLoop is not well-ordered. tag-20210426-1 4363 */ 4364 if( isOrderDistinct ){ 4365 if( iColumn>=0 4366 && j>=pLoop->u.btree.nEq 4367 && pIndex->pTable->aCol[iColumn].notNull==0 4368 ){ 4369 isOrderDistinct = 0; 4370 } 4371 if( iColumn==XN_EXPR ){ 4372 isOrderDistinct = 0; 4373 } 4374 } 4375 4376 /* Find the ORDER BY term that corresponds to the j-th column 4377 ** of the index and mark that ORDER BY term off 4378 */ 4379 isMatch = 0; 4380 for(i=0; bOnce && i<nOrderBy; i++){ 4381 if( MASKBIT(i) & obSat ) continue; 4382 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr); 4383 testcase( wctrlFlags & WHERE_GROUPBY ); 4384 testcase( wctrlFlags & WHERE_DISTINCTBY ); 4385 if( NEVER(pOBExpr==0) ) continue; 4386 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 4387 if( iColumn>=XN_ROWID ){ 4388 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue; 4389 if( pOBExpr->iTable!=iCur ) continue; 4390 if( pOBExpr->iColumn!=iColumn ) continue; 4391 }else{ 4392 Expr *pIdxExpr = pIndex->aColExpr->a[j].pExpr; 4393 if( sqlite3ExprCompareSkip(pOBExpr, pIdxExpr, iCur) ){ 4394 continue; 4395 } 4396 } 4397 if( iColumn!=XN_ROWID ){ 4398 pColl = sqlite3ExprNNCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 4399 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 4400 } 4401 if( wctrlFlags & WHERE_DISTINCTBY ){ 4402 pLoop->u.btree.nDistinctCol = j+1; 4403 } 4404 isMatch = 1; 4405 break; 4406 } 4407 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 4408 /* Make sure the sort order is compatible in an ORDER BY clause. 4409 ** Sort order is irrelevant for a GROUP BY clause. */ 4410 if( revSet ){ 4411 if( (rev ^ revIdx)!=(pOrderBy->a[i].sortFlags&KEYINFO_ORDER_DESC) ){ 4412 isMatch = 0; 4413 } 4414 }else{ 4415 rev = revIdx ^ (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC); 4416 if( rev ) *pRevMask |= MASKBIT(iLoop); 4417 revSet = 1; 4418 } 4419 } 4420 if( isMatch && (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL) ){ 4421 if( j==pLoop->u.btree.nEq ){ 4422 pLoop->wsFlags |= WHERE_BIGNULL_SORT; 4423 }else{ 4424 isMatch = 0; 4425 } 4426 } 4427 if( isMatch ){ 4428 if( iColumn==XN_ROWID ){ 4429 testcase( distinctColumns==0 ); 4430 distinctColumns = 1; 4431 } 4432 obSat |= MASKBIT(i); 4433 }else{ 4434 /* No match found */ 4435 if( j==0 || j<nKeyCol ){ 4436 testcase( isOrderDistinct!=0 ); 4437 isOrderDistinct = 0; 4438 } 4439 break; 4440 } 4441 } /* end Loop over all index columns */ 4442 if( distinctColumns ){ 4443 testcase( isOrderDistinct==0 ); 4444 isOrderDistinct = 1; 4445 } 4446 } /* end-if not one-row */ 4447 4448 /* Mark off any other ORDER BY terms that reference pLoop */ 4449 if( isOrderDistinct ){ 4450 orderDistinctMask |= pLoop->maskSelf; 4451 for(i=0; i<nOrderBy; i++){ 4452 Expr *p; 4453 Bitmask mTerm; 4454 if( MASKBIT(i) & obSat ) continue; 4455 p = pOrderBy->a[i].pExpr; 4456 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 4457 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 4458 if( (mTerm&~orderDistinctMask)==0 ){ 4459 obSat |= MASKBIT(i); 4460 } 4461 } 4462 } 4463 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 4464 if( obSat==obDone ) return (i8)nOrderBy; 4465 if( !isOrderDistinct ){ 4466 for(i=nOrderBy-1; i>0; i--){ 4467 Bitmask m = ALWAYS(i<BMS) ? MASKBIT(i) - 1 : 0; 4468 if( (obSat&m)==m ) return i; 4469 } 4470 return 0; 4471 } 4472 return -1; 4473 } 4474 4475 4476 /* 4477 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 4478 ** the planner assumes that the specified pOrderBy list is actually a GROUP 4479 ** BY clause - and so any order that groups rows as required satisfies the 4480 ** request. 4481 ** 4482 ** Normally, in this case it is not possible for the caller to determine 4483 ** whether or not the rows are really being delivered in sorted order, or 4484 ** just in some other order that provides the required grouping. However, 4485 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 4486 ** this function may be called on the returned WhereInfo object. It returns 4487 ** true if the rows really will be sorted in the specified order, or false 4488 ** otherwise. 4489 ** 4490 ** For example, assuming: 4491 ** 4492 ** CREATE INDEX i1 ON t1(x, Y); 4493 ** 4494 ** then 4495 ** 4496 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 4497 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 4498 */ 4499 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 4500 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 4501 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 4502 return pWInfo->sorted; 4503 } 4504 4505 #ifdef WHERETRACE_ENABLED 4506 /* For debugging use only: */ 4507 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 4508 static char zName[65]; 4509 int i; 4510 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 4511 if( pLast ) zName[i++] = pLast->cId; 4512 zName[i] = 0; 4513 return zName; 4514 } 4515 #endif 4516 4517 /* 4518 ** Return the cost of sorting nRow rows, assuming that the keys have 4519 ** nOrderby columns and that the first nSorted columns are already in 4520 ** order. 4521 */ 4522 static LogEst whereSortingCost( 4523 WhereInfo *pWInfo, 4524 LogEst nRow, 4525 int nOrderBy, 4526 int nSorted 4527 ){ 4528 /* TUNING: Estimated cost of a full external sort, where N is 4529 ** the number of rows to sort is: 4530 ** 4531 ** cost = (3.0 * N * log(N)). 4532 ** 4533 ** Or, if the order-by clause has X terms but only the last Y 4534 ** terms are out of order, then block-sorting will reduce the 4535 ** sorting cost to: 4536 ** 4537 ** cost = (3.0 * N * log(N)) * (Y/X) 4538 ** 4539 ** The (Y/X) term is implemented using stack variable rScale 4540 ** below. 4541 */ 4542 LogEst rScale, rSortCost; 4543 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 4544 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 4545 rSortCost = nRow + rScale + 16; 4546 4547 /* Multiple by log(M) where M is the number of output rows. 4548 ** Use the LIMIT for M if it is smaller. Or if this sort is for 4549 ** a DISTINCT operator, M will be the number of distinct output 4550 ** rows, so fudge it downwards a bit. 4551 */ 4552 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 4553 nRow = pWInfo->iLimit; 4554 }else if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT) ){ 4555 /* TUNING: In the sort for a DISTINCT operator, assume that the DISTINCT 4556 ** reduces the number of output rows by a factor of 2 */ 4557 if( nRow>10 ){ nRow -= 10; assert( 10==sqlite3LogEst(2) ); } 4558 } 4559 rSortCost += estLog(nRow); 4560 return rSortCost; 4561 } 4562 4563 /* 4564 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 4565 ** attempts to find the lowest cost path that visits each WhereLoop 4566 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 4567 ** 4568 ** Assume that the total number of output rows that will need to be sorted 4569 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 4570 ** costs if nRowEst==0. 4571 ** 4572 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 4573 ** error occurs. 4574 */ 4575 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 4576 int mxChoice; /* Maximum number of simultaneous paths tracked */ 4577 int nLoop; /* Number of terms in the join */ 4578 Parse *pParse; /* Parsing context */ 4579 sqlite3 *db; /* The database connection */ 4580 int iLoop; /* Loop counter over the terms of the join */ 4581 int ii, jj; /* Loop counters */ 4582 int mxI = 0; /* Index of next entry to replace */ 4583 int nOrderBy; /* Number of ORDER BY clause terms */ 4584 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 4585 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 4586 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 4587 WherePath *aFrom; /* All nFrom paths at the previous level */ 4588 WherePath *aTo; /* The nTo best paths at the current level */ 4589 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 4590 WherePath *pTo; /* An element of aTo[] that we are working on */ 4591 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 4592 WhereLoop **pX; /* Used to divy up the pSpace memory */ 4593 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 4594 char *pSpace; /* Temporary memory used by this routine */ 4595 int nSpace; /* Bytes of space allocated at pSpace */ 4596 4597 pParse = pWInfo->pParse; 4598 db = pParse->db; 4599 nLoop = pWInfo->nLevel; 4600 /* TUNING: For simple queries, only the best path is tracked. 4601 ** For 2-way joins, the 5 best paths are followed. 4602 ** For joins of 3 or more tables, track the 10 best paths */ 4603 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 4604 assert( nLoop<=pWInfo->pTabList->nSrc ); 4605 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 4606 4607 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 4608 ** case the purpose of this call is to estimate the number of rows returned 4609 ** by the overall query. Once this estimate has been obtained, the caller 4610 ** will invoke this function a second time, passing the estimate as the 4611 ** nRowEst parameter. */ 4612 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 4613 nOrderBy = 0; 4614 }else{ 4615 nOrderBy = pWInfo->pOrderBy->nExpr; 4616 } 4617 4618 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 4619 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 4620 nSpace += sizeof(LogEst) * nOrderBy; 4621 pSpace = sqlite3DbMallocRawNN(db, nSpace); 4622 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 4623 aTo = (WherePath*)pSpace; 4624 aFrom = aTo+mxChoice; 4625 memset(aFrom, 0, sizeof(aFrom[0])); 4626 pX = (WhereLoop**)(aFrom+mxChoice); 4627 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 4628 pFrom->aLoop = pX; 4629 } 4630 if( nOrderBy ){ 4631 /* If there is an ORDER BY clause and it is not being ignored, set up 4632 ** space for the aSortCost[] array. Each element of the aSortCost array 4633 ** is either zero - meaning it has not yet been initialized - or the 4634 ** cost of sorting nRowEst rows of data where the first X terms of 4635 ** the ORDER BY clause are already in order, where X is the array 4636 ** index. */ 4637 aSortCost = (LogEst*)pX; 4638 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 4639 } 4640 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 4641 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 4642 4643 /* Seed the search with a single WherePath containing zero WhereLoops. 4644 ** 4645 ** TUNING: Do not let the number of iterations go above 28. If the cost 4646 ** of computing an automatic index is not paid back within the first 28 4647 ** rows, then do not use the automatic index. */ 4648 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 4649 nFrom = 1; 4650 assert( aFrom[0].isOrdered==0 ); 4651 if( nOrderBy ){ 4652 /* If nLoop is zero, then there are no FROM terms in the query. Since 4653 ** in this case the query may return a maximum of one row, the results 4654 ** are already in the requested order. Set isOrdered to nOrderBy to 4655 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 4656 ** -1, indicating that the result set may or may not be ordered, 4657 ** depending on the loops added to the current plan. */ 4658 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 4659 } 4660 4661 /* Compute successively longer WherePaths using the previous generation 4662 ** of WherePaths as the basis for the next. Keep track of the mxChoice 4663 ** best paths at each generation */ 4664 for(iLoop=0; iLoop<nLoop; iLoop++){ 4665 nTo = 0; 4666 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 4667 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 4668 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 4669 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 4670 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 4671 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 4672 Bitmask maskNew; /* Mask of src visited by (..) */ 4673 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 4674 4675 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 4676 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 4677 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<3 ){ 4678 /* Do not use an automatic index if the this loop is expected 4679 ** to run less than 1.25 times. It is tempting to also exclude 4680 ** automatic index usage on an outer loop, but sometimes an automatic 4681 ** index is useful in the outer loop of a correlated subquery. */ 4682 assert( 10==sqlite3LogEst(2) ); 4683 continue; 4684 } 4685 4686 /* At this point, pWLoop is a candidate to be the next loop. 4687 ** Compute its cost */ 4688 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 4689 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 4690 nOut = pFrom->nRow + pWLoop->nOut; 4691 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 4692 if( isOrdered<0 ){ 4693 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 4694 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 4695 iLoop, pWLoop, &revMask); 4696 }else{ 4697 revMask = pFrom->revLoop; 4698 } 4699 if( isOrdered>=0 && isOrdered<nOrderBy ){ 4700 if( aSortCost[isOrdered]==0 ){ 4701 aSortCost[isOrdered] = whereSortingCost( 4702 pWInfo, nRowEst, nOrderBy, isOrdered 4703 ); 4704 } 4705 /* TUNING: Add a small extra penalty (5) to sorting as an 4706 ** extra encouragment to the query planner to select a plan 4707 ** where the rows emerge in the correct order without any sorting 4708 ** required. */ 4709 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]) + 5; 4710 4711 WHERETRACE(0x002, 4712 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 4713 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 4714 rUnsorted, rCost)); 4715 }else{ 4716 rCost = rUnsorted; 4717 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 4718 } 4719 4720 /* Check to see if pWLoop should be added to the set of 4721 ** mxChoice best-so-far paths. 4722 ** 4723 ** First look for an existing path among best-so-far paths 4724 ** that covers the same set of loops and has the same isOrdered 4725 ** setting as the current path candidate. 4726 ** 4727 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 4728 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 4729 ** of legal values for isOrdered, -1..64. 4730 */ 4731 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4732 if( pTo->maskLoop==maskNew 4733 && ((pTo->isOrdered^isOrdered)&0x80)==0 4734 ){ 4735 testcase( jj==nTo-1 ); 4736 break; 4737 } 4738 } 4739 if( jj>=nTo ){ 4740 /* None of the existing best-so-far paths match the candidate. */ 4741 if( nTo>=mxChoice 4742 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4743 ){ 4744 /* The current candidate is no better than any of the mxChoice 4745 ** paths currently in the best-so-far buffer. So discard 4746 ** this candidate as not viable. */ 4747 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4748 if( sqlite3WhereTrace&0x4 ){ 4749 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4750 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4751 isOrdered>=0 ? isOrdered+'0' : '?'); 4752 } 4753 #endif 4754 continue; 4755 } 4756 /* If we reach this points it means that the new candidate path 4757 ** needs to be added to the set of best-so-far paths. */ 4758 if( nTo<mxChoice ){ 4759 /* Increase the size of the aTo set by one */ 4760 jj = nTo++; 4761 }else{ 4762 /* New path replaces the prior worst to keep count below mxChoice */ 4763 jj = mxI; 4764 } 4765 pTo = &aTo[jj]; 4766 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4767 if( sqlite3WhereTrace&0x4 ){ 4768 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4769 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4770 isOrdered>=0 ? isOrdered+'0' : '?'); 4771 } 4772 #endif 4773 }else{ 4774 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4775 ** same set of loops and has the same isOrdered setting as the 4776 ** candidate path. Check to see if the candidate should replace 4777 ** pTo or if the candidate should be skipped. 4778 ** 4779 ** The conditional is an expanded vector comparison equivalent to: 4780 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4781 */ 4782 if( pTo->rCost<rCost 4783 || (pTo->rCost==rCost 4784 && (pTo->nRow<nOut 4785 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4786 ) 4787 ) 4788 ){ 4789 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4790 if( sqlite3WhereTrace&0x4 ){ 4791 sqlite3DebugPrintf( 4792 "Skip %s cost=%-3d,%3d,%3d order=%c", 4793 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4794 isOrdered>=0 ? isOrdered+'0' : '?'); 4795 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4796 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4797 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4798 } 4799 #endif 4800 /* Discard the candidate path from further consideration */ 4801 testcase( pTo->rCost==rCost ); 4802 continue; 4803 } 4804 testcase( pTo->rCost==rCost+1 ); 4805 /* Control reaches here if the candidate path is better than the 4806 ** pTo path. Replace pTo with the candidate. */ 4807 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4808 if( sqlite3WhereTrace&0x4 ){ 4809 sqlite3DebugPrintf( 4810 "Update %s cost=%-3d,%3d,%3d order=%c", 4811 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4812 isOrdered>=0 ? isOrdered+'0' : '?'); 4813 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4814 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4815 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4816 } 4817 #endif 4818 } 4819 /* pWLoop is a winner. Add it to the set of best so far */ 4820 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4821 pTo->revLoop = revMask; 4822 pTo->nRow = nOut; 4823 pTo->rCost = rCost; 4824 pTo->rUnsorted = rUnsorted; 4825 pTo->isOrdered = isOrdered; 4826 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4827 pTo->aLoop[iLoop] = pWLoop; 4828 if( nTo>=mxChoice ){ 4829 mxI = 0; 4830 mxCost = aTo[0].rCost; 4831 mxUnsorted = aTo[0].nRow; 4832 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4833 if( pTo->rCost>mxCost 4834 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4835 ){ 4836 mxCost = pTo->rCost; 4837 mxUnsorted = pTo->rUnsorted; 4838 mxI = jj; 4839 } 4840 } 4841 } 4842 } 4843 } 4844 4845 #ifdef WHERETRACE_ENABLED /* >=2 */ 4846 if( sqlite3WhereTrace & 0x02 ){ 4847 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4848 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4849 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4850 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4851 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4852 if( pTo->isOrdered>0 ){ 4853 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4854 }else{ 4855 sqlite3DebugPrintf("\n"); 4856 } 4857 } 4858 } 4859 #endif 4860 4861 /* Swap the roles of aFrom and aTo for the next generation */ 4862 pFrom = aTo; 4863 aTo = aFrom; 4864 aFrom = pFrom; 4865 nFrom = nTo; 4866 } 4867 4868 if( nFrom==0 ){ 4869 sqlite3ErrorMsg(pParse, "no query solution"); 4870 sqlite3DbFreeNN(db, pSpace); 4871 return SQLITE_ERROR; 4872 } 4873 4874 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4875 pFrom = aFrom; 4876 for(ii=1; ii<nFrom; ii++){ 4877 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4878 } 4879 assert( pWInfo->nLevel==nLoop ); 4880 /* Load the lowest cost path into pWInfo */ 4881 for(iLoop=0; iLoop<nLoop; iLoop++){ 4882 WhereLevel *pLevel = pWInfo->a + iLoop; 4883 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4884 pLevel->iFrom = pWLoop->iTab; 4885 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4886 } 4887 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4888 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4889 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4890 && nRowEst 4891 ){ 4892 Bitmask notUsed; 4893 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4894 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4895 if( rc==pWInfo->pResultSet->nExpr ){ 4896 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4897 } 4898 } 4899 pWInfo->bOrderedInnerLoop = 0; 4900 if( pWInfo->pOrderBy ){ 4901 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4902 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4903 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4904 } 4905 }else{ 4906 pWInfo->nOBSat = pFrom->isOrdered; 4907 pWInfo->revMask = pFrom->revLoop; 4908 if( pWInfo->nOBSat<=0 ){ 4909 pWInfo->nOBSat = 0; 4910 if( nLoop>0 ){ 4911 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4912 if( (wsFlags & WHERE_ONEROW)==0 4913 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4914 ){ 4915 Bitmask m = 0; 4916 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4917 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4918 testcase( wsFlags & WHERE_IPK ); 4919 testcase( wsFlags & WHERE_COLUMN_IN ); 4920 if( rc==pWInfo->pOrderBy->nExpr ){ 4921 pWInfo->bOrderedInnerLoop = 1; 4922 pWInfo->revMask = m; 4923 } 4924 } 4925 } 4926 }else if( nLoop 4927 && pWInfo->nOBSat==1 4928 && (pWInfo->wctrlFlags & (WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX))!=0 4929 ){ 4930 pWInfo->bOrderedInnerLoop = 1; 4931 } 4932 } 4933 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4934 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4935 ){ 4936 Bitmask revMask = 0; 4937 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4938 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4939 ); 4940 assert( pWInfo->sorted==0 ); 4941 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4942 pWInfo->sorted = 1; 4943 pWInfo->revMask = revMask; 4944 } 4945 } 4946 } 4947 4948 4949 pWInfo->nRowOut = pFrom->nRow; 4950 4951 /* Free temporary memory and return success */ 4952 sqlite3DbFreeNN(db, pSpace); 4953 return SQLITE_OK; 4954 } 4955 4956 /* 4957 ** Most queries use only a single table (they are not joins) and have 4958 ** simple == constraints against indexed fields. This routine attempts 4959 ** to plan those simple cases using much less ceremony than the 4960 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4961 ** times for the common case. 4962 ** 4963 ** Return non-zero on success, if this query can be handled by this 4964 ** no-frills query planner. Return zero if this query needs the 4965 ** general-purpose query planner. 4966 */ 4967 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4968 WhereInfo *pWInfo; 4969 SrcItem *pItem; 4970 WhereClause *pWC; 4971 WhereTerm *pTerm; 4972 WhereLoop *pLoop; 4973 int iCur; 4974 int j; 4975 Table *pTab; 4976 Index *pIdx; 4977 WhereScan scan; 4978 4979 pWInfo = pBuilder->pWInfo; 4980 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4981 assert( pWInfo->pTabList->nSrc>=1 ); 4982 pItem = pWInfo->pTabList->a; 4983 pTab = pItem->pTab; 4984 if( IsVirtual(pTab) ) return 0; 4985 if( pItem->fg.isIndexedBy ) return 0; 4986 iCur = pItem->iCursor; 4987 pWC = &pWInfo->sWC; 4988 pLoop = pBuilder->pNew; 4989 pLoop->wsFlags = 0; 4990 pLoop->nSkip = 0; 4991 pTerm = whereScanInit(&scan, pWC, iCur, -1, WO_EQ|WO_IS, 0); 4992 while( pTerm && pTerm->prereqRight ) pTerm = whereScanNext(&scan); 4993 if( pTerm ){ 4994 testcase( pTerm->eOperator & WO_IS ); 4995 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4996 pLoop->aLTerm[0] = pTerm; 4997 pLoop->nLTerm = 1; 4998 pLoop->u.btree.nEq = 1; 4999 /* TUNING: Cost of a rowid lookup is 10 */ 5000 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 5001 }else{ 5002 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 5003 int opMask; 5004 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 5005 if( !IsUniqueIndex(pIdx) 5006 || pIdx->pPartIdxWhere!=0 5007 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 5008 ) continue; 5009 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 5010 for(j=0; j<pIdx->nKeyCol; j++){ 5011 pTerm = whereScanInit(&scan, pWC, iCur, j, opMask, pIdx); 5012 while( pTerm && pTerm->prereqRight ) pTerm = whereScanNext(&scan); 5013 if( pTerm==0 ) break; 5014 testcase( pTerm->eOperator & WO_IS ); 5015 pLoop->aLTerm[j] = pTerm; 5016 } 5017 if( j!=pIdx->nKeyCol ) continue; 5018 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 5019 if( pIdx->isCovering || (pItem->colUsed & pIdx->colNotIdxed)==0 ){ 5020 pLoop->wsFlags |= WHERE_IDX_ONLY; 5021 } 5022 pLoop->nLTerm = j; 5023 pLoop->u.btree.nEq = j; 5024 pLoop->u.btree.pIndex = pIdx; 5025 /* TUNING: Cost of a unique index lookup is 15 */ 5026 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 5027 break; 5028 } 5029 } 5030 if( pLoop->wsFlags ){ 5031 pLoop->nOut = (LogEst)1; 5032 pWInfo->a[0].pWLoop = pLoop; 5033 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 5034 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 5035 pWInfo->a[0].iTabCur = iCur; 5036 pWInfo->nRowOut = 1; 5037 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 5038 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 5039 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 5040 } 5041 if( scan.iEquiv>1 ) pLoop->wsFlags |= WHERE_TRANSCONS; 5042 #ifdef SQLITE_DEBUG 5043 pLoop->cId = '0'; 5044 #endif 5045 #ifdef WHERETRACE_ENABLED 5046 if( sqlite3WhereTrace ){ 5047 sqlite3DebugPrintf("whereShortCut() used to compute solution\n"); 5048 } 5049 #endif 5050 return 1; 5051 } 5052 return 0; 5053 } 5054 5055 /* 5056 ** Helper function for exprIsDeterministic(). 5057 */ 5058 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 5059 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 5060 pWalker->eCode = 0; 5061 return WRC_Abort; 5062 } 5063 return WRC_Continue; 5064 } 5065 5066 /* 5067 ** Return true if the expression contains no non-deterministic SQL 5068 ** functions. Do not consider non-deterministic SQL functions that are 5069 ** part of sub-select statements. 5070 */ 5071 static int exprIsDeterministic(Expr *p){ 5072 Walker w; 5073 memset(&w, 0, sizeof(w)); 5074 w.eCode = 1; 5075 w.xExprCallback = exprNodeIsDeterministic; 5076 w.xSelectCallback = sqlite3SelectWalkFail; 5077 sqlite3WalkExpr(&w, p); 5078 return w.eCode; 5079 } 5080 5081 5082 #ifdef WHERETRACE_ENABLED 5083 /* 5084 ** Display all WhereLoops in pWInfo 5085 */ 5086 static void showAllWhereLoops(WhereInfo *pWInfo, WhereClause *pWC){ 5087 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 5088 WhereLoop *p; 5089 int i; 5090 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 5091 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 5092 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 5093 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 5094 sqlite3WhereLoopPrint(p, pWC); 5095 } 5096 } 5097 } 5098 # define WHERETRACE_ALL_LOOPS(W,C) showAllWhereLoops(W,C) 5099 #else 5100 # define WHERETRACE_ALL_LOOPS(W,C) 5101 #endif 5102 5103 /* Attempt to omit tables from a join that do not affect the result. 5104 ** For a table to not affect the result, the following must be true: 5105 ** 5106 ** 1) The query must not be an aggregate. 5107 ** 2) The table must be the RHS of a LEFT JOIN. 5108 ** 3) Either the query must be DISTINCT, or else the ON or USING clause 5109 ** must contain a constraint that limits the scan of the table to 5110 ** at most a single row. 5111 ** 4) The table must not be referenced by any part of the query apart 5112 ** from its own USING or ON clause. 5113 ** 5114 ** For example, given: 5115 ** 5116 ** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1); 5117 ** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2); 5118 ** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3); 5119 ** 5120 ** then table t2 can be omitted from the following: 5121 ** 5122 ** SELECT v1, v3 FROM t1 5123 ** LEFT JOIN t2 ON (t1.ipk=t2.ipk) 5124 ** LEFT JOIN t3 ON (t1.ipk=t3.ipk) 5125 ** 5126 ** or from: 5127 ** 5128 ** SELECT DISTINCT v1, v3 FROM t1 5129 ** LEFT JOIN t2 5130 ** LEFT JOIN t3 ON (t1.ipk=t3.ipk) 5131 */ 5132 static SQLITE_NOINLINE Bitmask whereOmitNoopJoin( 5133 WhereInfo *pWInfo, 5134 Bitmask notReady 5135 ){ 5136 int i; 5137 Bitmask tabUsed; 5138 5139 /* Preconditions checked by the caller */ 5140 assert( pWInfo->nLevel>=2 ); 5141 assert( OptimizationEnabled(pWInfo->pParse->db, SQLITE_OmitNoopJoin) ); 5142 5143 /* These two preconditions checked by the caller combine to guarantee 5144 ** condition (1) of the header comment */ 5145 assert( pWInfo->pResultSet!=0 ); 5146 assert( 0==(pWInfo->wctrlFlags & WHERE_AGG_DISTINCT) ); 5147 5148 tabUsed = sqlite3WhereExprListUsage(&pWInfo->sMaskSet, pWInfo->pResultSet); 5149 if( pWInfo->pOrderBy ){ 5150 tabUsed |= sqlite3WhereExprListUsage(&pWInfo->sMaskSet, pWInfo->pOrderBy); 5151 } 5152 for(i=pWInfo->nLevel-1; i>=1; i--){ 5153 WhereTerm *pTerm, *pEnd; 5154 SrcItem *pItem; 5155 WhereLoop *pLoop; 5156 pLoop = pWInfo->a[i].pWLoop; 5157 pItem = &pWInfo->pTabList->a[pLoop->iTab]; 5158 if( (pItem->fg.jointype & JT_LEFT)==0 ) continue; 5159 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)==0 5160 && (pLoop->wsFlags & WHERE_ONEROW)==0 5161 ){ 5162 continue; 5163 } 5164 if( (tabUsed & pLoop->maskSelf)!=0 ) continue; 5165 pEnd = pWInfo->sWC.a + pWInfo->sWC.nTerm; 5166 for(pTerm=pWInfo->sWC.a; pTerm<pEnd; pTerm++){ 5167 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 5168 if( !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 5169 || pTerm->pExpr->w.iRightJoinTable!=pItem->iCursor 5170 ){ 5171 break; 5172 } 5173 } 5174 } 5175 if( pTerm<pEnd ) continue; 5176 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 5177 notReady &= ~pLoop->maskSelf; 5178 for(pTerm=pWInfo->sWC.a; pTerm<pEnd; pTerm++){ 5179 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 5180 pTerm->wtFlags |= TERM_CODED; 5181 } 5182 } 5183 if( i!=pWInfo->nLevel-1 ){ 5184 int nByte = (pWInfo->nLevel-1-i) * sizeof(WhereLevel); 5185 memmove(&pWInfo->a[i], &pWInfo->a[i+1], nByte); 5186 } 5187 pWInfo->nLevel--; 5188 assert( pWInfo->nLevel>0 ); 5189 } 5190 return notReady; 5191 } 5192 5193 /* 5194 ** Check to see if there are any SEARCH loops that might benefit from 5195 ** using a Bloom filter. Consider a Bloom filter if: 5196 ** 5197 ** (1) The SEARCH happens more than N times where N is the number 5198 ** of rows in the table that is being considered for the Bloom 5199 ** filter. 5200 ** (2) Some searches are expected to find zero rows. (This is determined 5201 ** by the WHERE_SELFCULL flag on the term.) 5202 ** (3) Bloom-filter processing is not disabled. (Checked by the 5203 ** caller.) 5204 ** (4) The size of the table being searched is known by ANALYZE. 5205 ** 5206 ** This block of code merely checks to see if a Bloom filter would be 5207 ** appropriate, and if so sets the WHERE_BLOOMFILTER flag on the 5208 ** WhereLoop. The implementation of the Bloom filter comes further 5209 ** down where the code for each WhereLoop is generated. 5210 */ 5211 static SQLITE_NOINLINE void whereCheckIfBloomFilterIsUseful( 5212 const WhereInfo *pWInfo 5213 ){ 5214 int i; 5215 LogEst nSearch; 5216 5217 assert( pWInfo->nLevel>=2 ); 5218 assert( OptimizationEnabled(pWInfo->pParse->db, SQLITE_BloomFilter) ); 5219 nSearch = pWInfo->a[0].pWLoop->nOut; 5220 for(i=1; i<pWInfo->nLevel; i++){ 5221 WhereLoop *pLoop = pWInfo->a[i].pWLoop; 5222 const unsigned int reqFlags = (WHERE_SELFCULL|WHERE_COLUMN_EQ); 5223 if( (pLoop->wsFlags & reqFlags)==reqFlags 5224 /* vvvvvv--- Always the case if WHERE_COLUMN_EQ is defined */ 5225 && ALWAYS((pLoop->wsFlags & (WHERE_IPK|WHERE_INDEXED))!=0) 5226 ){ 5227 SrcItem *pItem = &pWInfo->pTabList->a[pLoop->iTab]; 5228 Table *pTab = pItem->pTab; 5229 pTab->tabFlags |= TF_StatsUsed; 5230 if( nSearch > pTab->nRowLogEst 5231 && (pTab->tabFlags & TF_HasStat1)!=0 5232 ){ 5233 testcase( pItem->fg.jointype & JT_LEFT ); 5234 pLoop->wsFlags |= WHERE_BLOOMFILTER; 5235 pLoop->wsFlags &= ~WHERE_IDX_ONLY; 5236 WHERETRACE(0xffff, ( 5237 "-> use Bloom-filter on loop %c because there are ~%.1e " 5238 "lookups into %s which has only ~%.1e rows\n", 5239 pLoop->cId, (double)sqlite3LogEstToInt(nSearch), pTab->zName, 5240 (double)sqlite3LogEstToInt(pTab->nRowLogEst))); 5241 } 5242 } 5243 nSearch += pLoop->nOut; 5244 } 5245 } 5246 5247 /* 5248 ** Generate the beginning of the loop used for WHERE clause processing. 5249 ** The return value is a pointer to an opaque structure that contains 5250 ** information needed to terminate the loop. Later, the calling routine 5251 ** should invoke sqlite3WhereEnd() with the return value of this function 5252 ** in order to complete the WHERE clause processing. 5253 ** 5254 ** If an error occurs, this routine returns NULL. 5255 ** 5256 ** The basic idea is to do a nested loop, one loop for each table in 5257 ** the FROM clause of a select. (INSERT and UPDATE statements are the 5258 ** same as a SELECT with only a single table in the FROM clause.) For 5259 ** example, if the SQL is this: 5260 ** 5261 ** SELECT * FROM t1, t2, t3 WHERE ...; 5262 ** 5263 ** Then the code generated is conceptually like the following: 5264 ** 5265 ** foreach row1 in t1 do \ Code generated 5266 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 5267 ** foreach row3 in t3 do / 5268 ** ... 5269 ** end \ Code generated 5270 ** end |-- by sqlite3WhereEnd() 5271 ** end / 5272 ** 5273 ** Note that the loops might not be nested in the order in which they 5274 ** appear in the FROM clause if a different order is better able to make 5275 ** use of indices. Note also that when the IN operator appears in 5276 ** the WHERE clause, it might result in additional nested loops for 5277 ** scanning through all values on the right-hand side of the IN. 5278 ** 5279 ** There are Btree cursors associated with each table. t1 uses cursor 5280 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 5281 ** And so forth. This routine generates code to open those VDBE cursors 5282 ** and sqlite3WhereEnd() generates the code to close them. 5283 ** 5284 ** The code that sqlite3WhereBegin() generates leaves the cursors named 5285 ** in pTabList pointing at their appropriate entries. The [...] code 5286 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 5287 ** data from the various tables of the loop. 5288 ** 5289 ** If the WHERE clause is empty, the foreach loops must each scan their 5290 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 5291 ** the tables have indices and there are terms in the WHERE clause that 5292 ** refer to those indices, a complete table scan can be avoided and the 5293 ** code will run much faster. Most of the work of this routine is checking 5294 ** to see if there are indices that can be used to speed up the loop. 5295 ** 5296 ** Terms of the WHERE clause are also used to limit which rows actually 5297 ** make it to the "..." in the middle of the loop. After each "foreach", 5298 ** terms of the WHERE clause that use only terms in that loop and outer 5299 ** loops are evaluated and if false a jump is made around all subsequent 5300 ** inner loops (or around the "..." if the test occurs within the inner- 5301 ** most loop) 5302 ** 5303 ** OUTER JOINS 5304 ** 5305 ** An outer join of tables t1 and t2 is conceptally coded as follows: 5306 ** 5307 ** foreach row1 in t1 do 5308 ** flag = 0 5309 ** foreach row2 in t2 do 5310 ** start: 5311 ** ... 5312 ** flag = 1 5313 ** end 5314 ** if flag==0 then 5315 ** move the row2 cursor to a null row 5316 ** goto start 5317 ** fi 5318 ** end 5319 ** 5320 ** ORDER BY CLAUSE PROCESSING 5321 ** 5322 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 5323 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 5324 ** if there is one. If there is no ORDER BY clause or if this routine 5325 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 5326 ** 5327 ** The iIdxCur parameter is the cursor number of an index. If 5328 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 5329 ** to use for OR clause processing. The WHERE clause should use this 5330 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 5331 ** the first cursor in an array of cursors for all indices. iIdxCur should 5332 ** be used to compute the appropriate cursor depending on which index is 5333 ** used. 5334 */ 5335 WhereInfo *sqlite3WhereBegin( 5336 Parse *pParse, /* The parser context */ 5337 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 5338 Expr *pWhere, /* The WHERE clause */ 5339 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 5340 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 5341 Select *pLimit, /* Use this LIMIT/OFFSET clause, if any */ 5342 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 5343 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 5344 ** If WHERE_USE_LIMIT, then the limit amount */ 5345 ){ 5346 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 5347 int nTabList; /* Number of elements in pTabList */ 5348 WhereInfo *pWInfo; /* Will become the return value of this function */ 5349 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 5350 Bitmask notReady; /* Cursors that are not yet positioned */ 5351 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 5352 WhereMaskSet *pMaskSet; /* The expression mask set */ 5353 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 5354 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 5355 int ii; /* Loop counter */ 5356 sqlite3 *db; /* Database connection */ 5357 int rc; /* Return code */ 5358 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 5359 5360 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 5361 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 5362 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5363 )); 5364 5365 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 5366 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5367 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 5368 5369 /* Variable initialization */ 5370 db = pParse->db; 5371 memset(&sWLB, 0, sizeof(sWLB)); 5372 5373 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 5374 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 5375 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 5376 5377 /* The number of tables in the FROM clause is limited by the number of 5378 ** bits in a Bitmask 5379 */ 5380 testcase( pTabList->nSrc==BMS ); 5381 if( pTabList->nSrc>BMS ){ 5382 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 5383 return 0; 5384 } 5385 5386 /* This function normally generates a nested loop for all tables in 5387 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 5388 ** only generate code for the first table in pTabList and assume that 5389 ** any cursors associated with subsequent tables are uninitialized. 5390 */ 5391 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 5392 5393 /* Allocate and initialize the WhereInfo structure that will become the 5394 ** return value. A single allocation is used to store the WhereInfo 5395 ** struct, the contents of WhereInfo.a[], the WhereClause structure 5396 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 5397 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 5398 ** some architectures. Hence the ROUND8() below. 5399 */ 5400 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 5401 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 5402 if( db->mallocFailed ){ 5403 sqlite3DbFree(db, pWInfo); 5404 pWInfo = 0; 5405 goto whereBeginError; 5406 } 5407 pWInfo->pParse = pParse; 5408 pWInfo->pTabList = pTabList; 5409 pWInfo->pOrderBy = pOrderBy; 5410 pWInfo->pWhere = pWhere; 5411 pWInfo->pResultSet = pResultSet; 5412 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 5413 pWInfo->nLevel = nTabList; 5414 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(pParse); 5415 pWInfo->wctrlFlags = wctrlFlags; 5416 pWInfo->iLimit = iAuxArg; 5417 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 5418 #ifndef SQLITE_OMIT_VIRTUALTABLE 5419 pWInfo->pLimit = pLimit; 5420 #endif 5421 memset(&pWInfo->nOBSat, 0, 5422 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 5423 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 5424 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 5425 pMaskSet = &pWInfo->sMaskSet; 5426 pMaskSet->n = 0; 5427 pMaskSet->ix[0] = -99; /* Initialize ix[0] to a value that can never be 5428 ** a valid cursor number, to avoid an initial 5429 ** test for pMaskSet->n==0 in sqlite3WhereGetMask() */ 5430 sWLB.pWInfo = pWInfo; 5431 sWLB.pWC = &pWInfo->sWC; 5432 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 5433 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 5434 whereLoopInit(sWLB.pNew); 5435 #ifdef SQLITE_DEBUG 5436 sWLB.pNew->cId = '*'; 5437 #endif 5438 5439 /* Split the WHERE clause into separate subexpressions where each 5440 ** subexpression is separated by an AND operator. 5441 */ 5442 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 5443 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 5444 5445 /* Special case: No FROM clause 5446 */ 5447 if( nTabList==0 ){ 5448 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 5449 if( (wctrlFlags & WHERE_WANT_DISTINCT)!=0 5450 && OptimizationEnabled(db, SQLITE_DistinctOpt) 5451 ){ 5452 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 5453 } 5454 ExplainQueryPlan((pParse, 0, "SCAN CONSTANT ROW")); 5455 }else{ 5456 /* Assign a bit from the bitmask to every term in the FROM clause. 5457 ** 5458 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 5459 ** 5460 ** The rule of the previous sentence ensures thta if X is the bitmask for 5461 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 5462 ** Knowing the bitmask for all tables to the left of a left join is 5463 ** important. Ticket #3015. 5464 ** 5465 ** Note that bitmasks are created for all pTabList->nSrc tables in 5466 ** pTabList, not just the first nTabList tables. nTabList is normally 5467 ** equal to pTabList->nSrc but might be shortened to 1 if the 5468 ** WHERE_OR_SUBCLAUSE flag is set. 5469 */ 5470 ii = 0; 5471 do{ 5472 createMask(pMaskSet, pTabList->a[ii].iCursor); 5473 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 5474 }while( (++ii)<pTabList->nSrc ); 5475 #ifdef SQLITE_DEBUG 5476 { 5477 Bitmask mx = 0; 5478 for(ii=0; ii<pTabList->nSrc; ii++){ 5479 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 5480 assert( m>=mx ); 5481 mx = m; 5482 } 5483 } 5484 #endif 5485 } 5486 5487 /* Analyze all of the subexpressions. */ 5488 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 5489 sqlite3WhereAddLimit(&pWInfo->sWC, pLimit); 5490 if( db->mallocFailed ) goto whereBeginError; 5491 5492 /* Special case: WHERE terms that do not refer to any tables in the join 5493 ** (constant expressions). Evaluate each such term, and jump over all the 5494 ** generated code if the result is not true. 5495 ** 5496 ** Do not do this if the expression contains non-deterministic functions 5497 ** that are not within a sub-select. This is not strictly required, but 5498 ** preserves SQLite's legacy behaviour in the following two cases: 5499 ** 5500 ** FROM ... WHERE random()>0; -- eval random() once per row 5501 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 5502 */ 5503 for(ii=0; ii<sWLB.pWC->nBase; ii++){ 5504 WhereTerm *pT = &sWLB.pWC->a[ii]; 5505 if( pT->wtFlags & TERM_VIRTUAL ) continue; 5506 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 5507 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 5508 pT->wtFlags |= TERM_CODED; 5509 } 5510 } 5511 5512 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 5513 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 5514 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 5515 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 5516 wctrlFlags &= ~WHERE_WANT_DISTINCT; 5517 pWInfo->wctrlFlags &= ~WHERE_WANT_DISTINCT; 5518 }else if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 5519 /* The DISTINCT marking is pointless. Ignore it. */ 5520 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 5521 }else if( pOrderBy==0 ){ 5522 /* Try to ORDER BY the result set to make distinct processing easier */ 5523 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 5524 pWInfo->pOrderBy = pResultSet; 5525 } 5526 } 5527 5528 /* Construct the WhereLoop objects */ 5529 #if defined(WHERETRACE_ENABLED) 5530 if( sqlite3WhereTrace & 0xffff ){ 5531 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 5532 if( wctrlFlags & WHERE_USE_LIMIT ){ 5533 sqlite3DebugPrintf(", limit: %d", iAuxArg); 5534 } 5535 sqlite3DebugPrintf(")\n"); 5536 if( sqlite3WhereTrace & 0x100 ){ 5537 Select sSelect; 5538 memset(&sSelect, 0, sizeof(sSelect)); 5539 sSelect.selFlags = SF_WhereBegin; 5540 sSelect.pSrc = pTabList; 5541 sSelect.pWhere = pWhere; 5542 sSelect.pOrderBy = pOrderBy; 5543 sSelect.pEList = pResultSet; 5544 sqlite3TreeViewSelect(0, &sSelect, 0); 5545 } 5546 } 5547 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 5548 sqlite3DebugPrintf("---- WHERE clause at start of analysis:\n"); 5549 sqlite3WhereClausePrint(sWLB.pWC); 5550 } 5551 #endif 5552 5553 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 5554 rc = whereLoopAddAll(&sWLB); 5555 if( rc ) goto whereBeginError; 5556 5557 #ifdef SQLITE_ENABLE_STAT4 5558 /* If one or more WhereTerm.truthProb values were used in estimating 5559 ** loop parameters, but then those truthProb values were subsequently 5560 ** changed based on STAT4 information while computing subsequent loops, 5561 ** then we need to rerun the whole loop building process so that all 5562 ** loops will be built using the revised truthProb values. */ 5563 if( sWLB.bldFlags2 & SQLITE_BLDF2_2NDPASS ){ 5564 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC); 5565 WHERETRACE(0xffff, 5566 ("**** Redo all loop computations due to" 5567 " TERM_HIGHTRUTH changes ****\n")); 5568 while( pWInfo->pLoops ){ 5569 WhereLoop *p = pWInfo->pLoops; 5570 pWInfo->pLoops = p->pNextLoop; 5571 whereLoopDelete(db, p); 5572 } 5573 rc = whereLoopAddAll(&sWLB); 5574 if( rc ) goto whereBeginError; 5575 } 5576 #endif 5577 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC); 5578 5579 wherePathSolver(pWInfo, 0); 5580 if( db->mallocFailed ) goto whereBeginError; 5581 if( pWInfo->pOrderBy ){ 5582 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 5583 if( db->mallocFailed ) goto whereBeginError; 5584 } 5585 } 5586 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 5587 pWInfo->revMask = ALLBITS; 5588 } 5589 if( pParse->nErr ){ 5590 goto whereBeginError; 5591 } 5592 assert( db->mallocFailed==0 ); 5593 #ifdef WHERETRACE_ENABLED 5594 if( sqlite3WhereTrace ){ 5595 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 5596 if( pWInfo->nOBSat>0 ){ 5597 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 5598 } 5599 switch( pWInfo->eDistinct ){ 5600 case WHERE_DISTINCT_UNIQUE: { 5601 sqlite3DebugPrintf(" DISTINCT=unique"); 5602 break; 5603 } 5604 case WHERE_DISTINCT_ORDERED: { 5605 sqlite3DebugPrintf(" DISTINCT=ordered"); 5606 break; 5607 } 5608 case WHERE_DISTINCT_UNORDERED: { 5609 sqlite3DebugPrintf(" DISTINCT=unordered"); 5610 break; 5611 } 5612 } 5613 sqlite3DebugPrintf("\n"); 5614 for(ii=0; ii<pWInfo->nLevel; ii++){ 5615 sqlite3WhereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 5616 } 5617 } 5618 #endif 5619 5620 /* Attempt to omit tables from a join that do not affect the result. 5621 ** See the comment on whereOmitNoopJoin() for further information. 5622 ** 5623 ** This query optimization is factored out into a separate "no-inline" 5624 ** procedure to keep the sqlite3WhereBegin() procedure from becoming 5625 ** too large. If sqlite3WhereBegin() becomes too large, that prevents 5626 ** some C-compiler optimizers from in-lining the 5627 ** sqlite3WhereCodeOneLoopStart() procedure, and it is important to 5628 ** in-line sqlite3WhereCodeOneLoopStart() for performance reasons. 5629 */ 5630 notReady = ~(Bitmask)0; 5631 if( pWInfo->nLevel>=2 5632 && pResultSet!=0 /* these two combine to guarantee */ 5633 && 0==(wctrlFlags & WHERE_AGG_DISTINCT) /* condition (1) above */ 5634 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 5635 ){ 5636 notReady = whereOmitNoopJoin(pWInfo, notReady); 5637 nTabList = pWInfo->nLevel; 5638 assert( nTabList>0 ); 5639 } 5640 5641 /* Check to see if there are any SEARCH loops that might benefit from 5642 ** using a Bloom filter. 5643 */ 5644 if( pWInfo->nLevel>=2 5645 && OptimizationEnabled(db, SQLITE_BloomFilter) 5646 ){ 5647 whereCheckIfBloomFilterIsUseful(pWInfo); 5648 } 5649 5650 #if defined(WHERETRACE_ENABLED) 5651 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 5652 sqlite3DebugPrintf("---- WHERE clause at end of analysis:\n"); 5653 sqlite3WhereClausePrint(sWLB.pWC); 5654 } 5655 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 5656 #endif 5657 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 5658 5659 /* If the caller is an UPDATE or DELETE statement that is requesting 5660 ** to use a one-pass algorithm, determine if this is appropriate. 5661 ** 5662 ** A one-pass approach can be used if the caller has requested one 5663 ** and either (a) the scan visits at most one row or (b) each 5664 ** of the following are true: 5665 ** 5666 ** * the caller has indicated that a one-pass approach can be used 5667 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and 5668 ** * the table is not a virtual table, and 5669 ** * either the scan does not use the OR optimization or the caller 5670 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified 5671 ** for DELETE). 5672 ** 5673 ** The last qualification is because an UPDATE statement uses 5674 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can 5675 ** use a one-pass approach, and this is not set accurately for scans 5676 ** that use the OR optimization. 5677 */ 5678 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 5679 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 5680 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 5681 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 5682 assert( !(wsFlags & WHERE_VIRTUALTABLE) || IsVirtual(pTabList->a[0].pTab) ); 5683 if( bOnerow || ( 5684 0!=(wctrlFlags & WHERE_ONEPASS_MULTIROW) 5685 && !IsVirtual(pTabList->a[0].pTab) 5686 && (0==(wsFlags & WHERE_MULTI_OR) || (wctrlFlags & WHERE_DUPLICATES_OK)) 5687 )){ 5688 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 5689 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 5690 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 5691 bFordelete = OPFLAG_FORDELETE; 5692 } 5693 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 5694 } 5695 } 5696 } 5697 5698 /* Open all tables in the pTabList and any indices selected for 5699 ** searching those tables. 5700 */ 5701 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 5702 Table *pTab; /* Table to open */ 5703 int iDb; /* Index of database containing table/index */ 5704 SrcItem *pTabItem; 5705 5706 pTabItem = &pTabList->a[pLevel->iFrom]; 5707 pTab = pTabItem->pTab; 5708 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 5709 pLoop = pLevel->pWLoop; 5710 if( (pTab->tabFlags & TF_Ephemeral)!=0 || IsView(pTab) ){ 5711 /* Do nothing */ 5712 }else 5713 #ifndef SQLITE_OMIT_VIRTUALTABLE 5714 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 5715 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 5716 int iCur = pTabItem->iCursor; 5717 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 5718 }else if( IsVirtual(pTab) ){ 5719 /* noop */ 5720 }else 5721 #endif 5722 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 5723 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 5724 int op = OP_OpenRead; 5725 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5726 op = OP_OpenWrite; 5727 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 5728 }; 5729 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 5730 assert( pTabItem->iCursor==pLevel->iTabCur ); 5731 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 5732 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 5733 if( pWInfo->eOnePass==ONEPASS_OFF 5734 && pTab->nCol<BMS 5735 && (pTab->tabFlags & (TF_HasGenerated|TF_WithoutRowid))==0 5736 && (pLoop->wsFlags & (WHERE_AUTO_INDEX|WHERE_BLOOMFILTER))==0 5737 ){ 5738 /* If we know that only a prefix of the record will be used, 5739 ** it is advantageous to reduce the "column count" field in 5740 ** the P4 operand of the OP_OpenRead/Write opcode. */ 5741 Bitmask b = pTabItem->colUsed; 5742 int n = 0; 5743 for(; b; b=b>>1, n++){} 5744 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 5745 assert( n<=pTab->nCol ); 5746 } 5747 #ifdef SQLITE_ENABLE_CURSOR_HINTS 5748 if( pLoop->u.btree.pIndex!=0 ){ 5749 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 5750 }else 5751 #endif 5752 { 5753 sqlite3VdbeChangeP5(v, bFordelete); 5754 } 5755 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5756 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 5757 (const u8*)&pTabItem->colUsed, P4_INT64); 5758 #endif 5759 }else{ 5760 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 5761 } 5762 if( pLoop->wsFlags & WHERE_INDEXED ){ 5763 Index *pIx = pLoop->u.btree.pIndex; 5764 int iIndexCur; 5765 int op = OP_OpenRead; 5766 /* iAuxArg is always set to a positive value if ONEPASS is possible */ 5767 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 5768 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 5769 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 5770 ){ 5771 /* This is one term of an OR-optimization using the PRIMARY KEY of a 5772 ** WITHOUT ROWID table. No need for a separate index */ 5773 iIndexCur = pLevel->iTabCur; 5774 op = 0; 5775 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5776 Index *pJ = pTabItem->pTab->pIndex; 5777 iIndexCur = iAuxArg; 5778 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 5779 while( ALWAYS(pJ) && pJ!=pIx ){ 5780 iIndexCur++; 5781 pJ = pJ->pNext; 5782 } 5783 op = OP_OpenWrite; 5784 pWInfo->aiCurOnePass[1] = iIndexCur; 5785 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 5786 iIndexCur = iAuxArg; 5787 op = OP_ReopenIdx; 5788 }else{ 5789 iIndexCur = pParse->nTab++; 5790 } 5791 pLevel->iIdxCur = iIndexCur; 5792 assert( pIx->pSchema==pTab->pSchema ); 5793 assert( iIndexCur>=0 ); 5794 if( op ){ 5795 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 5796 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 5797 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 5798 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 5799 && (pLoop->wsFlags & WHERE_BIGNULL_SORT)==0 5800 && (pLoop->wsFlags & WHERE_IN_SEEKSCAN)==0 5801 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 5802 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 5803 ){ 5804 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); 5805 } 5806 VdbeComment((v, "%s", pIx->zName)); 5807 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5808 { 5809 u64 colUsed = 0; 5810 int ii, jj; 5811 for(ii=0; ii<pIx->nColumn; ii++){ 5812 jj = pIx->aiColumn[ii]; 5813 if( jj<0 ) continue; 5814 if( jj>63 ) jj = 63; 5815 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 5816 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 5817 } 5818 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 5819 (u8*)&colUsed, P4_INT64); 5820 } 5821 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 5822 } 5823 } 5824 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 5825 } 5826 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 5827 if( db->mallocFailed ) goto whereBeginError; 5828 5829 /* Generate the code to do the search. Each iteration of the for 5830 ** loop below generates code for a single nested loop of the VM 5831 ** program. 5832 */ 5833 for(ii=0; ii<nTabList; ii++){ 5834 int addrExplain; 5835 int wsFlags; 5836 if( pParse->nErr ) goto whereBeginError; 5837 pLevel = &pWInfo->a[ii]; 5838 wsFlags = pLevel->pWLoop->wsFlags; 5839 if( (wsFlags & (WHERE_AUTO_INDEX|WHERE_BLOOMFILTER))!=0 ){ 5840 if( (wsFlags & WHERE_AUTO_INDEX)!=0 ){ 5841 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 5842 constructAutomaticIndex(pParse, &pWInfo->sWC, 5843 &pTabList->a[pLevel->iFrom], notReady, pLevel); 5844 #endif 5845 }else{ 5846 sqlite3ConstructBloomFilter(pWInfo, ii, pLevel, notReady); 5847 } 5848 if( db->mallocFailed ) goto whereBeginError; 5849 } 5850 addrExplain = sqlite3WhereExplainOneScan( 5851 pParse, pTabList, pLevel, wctrlFlags 5852 ); 5853 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 5854 notReady = sqlite3WhereCodeOneLoopStart(pParse,v,pWInfo,ii,pLevel,notReady); 5855 pWInfo->iContinue = pLevel->addrCont; 5856 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 5857 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 5858 } 5859 } 5860 5861 /* Done. */ 5862 VdbeModuleComment((v, "Begin WHERE-core")); 5863 pWInfo->iEndWhere = sqlite3VdbeCurrentAddr(v); 5864 return pWInfo; 5865 5866 /* Jump here if malloc fails */ 5867 whereBeginError: 5868 if( pWInfo ){ 5869 testcase( pWInfo->pExprMods!=0 ); 5870 whereUndoExprMods(pWInfo); 5871 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5872 whereInfoFree(db, pWInfo); 5873 } 5874 return 0; 5875 } 5876 5877 /* 5878 ** Part of sqlite3WhereEnd() will rewrite opcodes to reference the 5879 ** index rather than the main table. In SQLITE_DEBUG mode, we want 5880 ** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine 5881 ** does that. 5882 */ 5883 #ifndef SQLITE_DEBUG 5884 # define OpcodeRewriteTrace(D,K,P) /* no-op */ 5885 #else 5886 # define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P) 5887 static void sqlite3WhereOpcodeRewriteTrace( 5888 sqlite3 *db, 5889 int pc, 5890 VdbeOp *pOp 5891 ){ 5892 if( (db->flags & SQLITE_VdbeAddopTrace)==0 ) return; 5893 sqlite3VdbePrintOp(0, pc, pOp); 5894 } 5895 #endif 5896 5897 /* 5898 ** Generate the end of the WHERE loop. See comments on 5899 ** sqlite3WhereBegin() for additional information. 5900 */ 5901 void sqlite3WhereEnd(WhereInfo *pWInfo){ 5902 Parse *pParse = pWInfo->pParse; 5903 Vdbe *v = pParse->pVdbe; 5904 int i; 5905 WhereLevel *pLevel; 5906 WhereLoop *pLoop; 5907 SrcList *pTabList = pWInfo->pTabList; 5908 sqlite3 *db = pParse->db; 5909 int iEnd = sqlite3VdbeCurrentAddr(v); 5910 5911 /* Generate loop termination code. 5912 */ 5913 VdbeModuleComment((v, "End WHERE-core")); 5914 for(i=pWInfo->nLevel-1; i>=0; i--){ 5915 int addr; 5916 pLevel = &pWInfo->a[i]; 5917 pLoop = pLevel->pWLoop; 5918 if( pLevel->op!=OP_Noop ){ 5919 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5920 int addrSeek = 0; 5921 Index *pIdx; 5922 int n; 5923 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 5924 && i==pWInfo->nLevel-1 /* Ticket [ef9318757b152e3] 2017-10-21 */ 5925 && (pLoop->wsFlags & WHERE_INDEXED)!=0 5926 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 5927 && (n = pLoop->u.btree.nDistinctCol)>0 5928 && pIdx->aiRowLogEst[n]>=36 5929 ){ 5930 int r1 = pParse->nMem+1; 5931 int j, op; 5932 for(j=0; j<n; j++){ 5933 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 5934 } 5935 pParse->nMem += n+1; 5936 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 5937 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 5938 VdbeCoverageIf(v, op==OP_SeekLT); 5939 VdbeCoverageIf(v, op==OP_SeekGT); 5940 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 5941 } 5942 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 5943 /* The common case: Advance to the next row */ 5944 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5945 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 5946 sqlite3VdbeChangeP5(v, pLevel->p5); 5947 VdbeCoverage(v); 5948 VdbeCoverageIf(v, pLevel->op==OP_Next); 5949 VdbeCoverageIf(v, pLevel->op==OP_Prev); 5950 VdbeCoverageIf(v, pLevel->op==OP_VNext); 5951 if( pLevel->regBignull ){ 5952 sqlite3VdbeResolveLabel(v, pLevel->addrBignull); 5953 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, pLevel->regBignull, pLevel->p2-1); 5954 VdbeCoverage(v); 5955 } 5956 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5957 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 5958 #endif 5959 }else{ 5960 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5961 } 5962 if( (pLoop->wsFlags & WHERE_IN_ABLE)!=0 && pLevel->u.in.nIn>0 ){ 5963 struct InLoop *pIn; 5964 int j; 5965 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 5966 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 5967 assert( sqlite3VdbeGetOp(v, pIn->addrInTop+1)->opcode==OP_IsNull 5968 || pParse->db->mallocFailed ); 5969 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5970 if( pIn->eEndLoopOp!=OP_Noop ){ 5971 if( pIn->nPrefix ){ 5972 int bEarlyOut = 5973 (pLoop->wsFlags & WHERE_VIRTUALTABLE)==0 5974 && (pLoop->wsFlags & WHERE_IN_EARLYOUT)!=0; 5975 if( pLevel->iLeftJoin ){ 5976 /* For LEFT JOIN queries, cursor pIn->iCur may not have been 5977 ** opened yet. This occurs for WHERE clauses such as 5978 ** "a = ? AND b IN (...)", where the index is on (a, b). If 5979 ** the RHS of the (a=?) is NULL, then the "b IN (...)" may 5980 ** never have been coded, but the body of the loop run to 5981 ** return the null-row. So, if the cursor is not open yet, 5982 ** jump over the OP_Next or OP_Prev instruction about to 5983 ** be coded. */ 5984 sqlite3VdbeAddOp2(v, OP_IfNotOpen, pIn->iCur, 5985 sqlite3VdbeCurrentAddr(v) + 2 + bEarlyOut); 5986 VdbeCoverage(v); 5987 } 5988 if( bEarlyOut ){ 5989 sqlite3VdbeAddOp4Int(v, OP_IfNoHope, pLevel->iIdxCur, 5990 sqlite3VdbeCurrentAddr(v)+2, 5991 pIn->iBase, pIn->nPrefix); 5992 VdbeCoverage(v); 5993 /* Retarget the OP_IsNull against the left operand of IN so 5994 ** it jumps past the OP_IfNoHope. This is because the 5995 ** OP_IsNull also bypasses the OP_Affinity opcode that is 5996 ** required by OP_IfNoHope. */ 5997 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5998 } 5999 } 6000 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 6001 VdbeCoverage(v); 6002 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Prev); 6003 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Next); 6004 } 6005 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 6006 } 6007 } 6008 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 6009 if( pLevel->addrSkip ){ 6010 sqlite3VdbeGoto(v, pLevel->addrSkip); 6011 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 6012 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 6013 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 6014 } 6015 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 6016 if( pLevel->addrLikeRep ){ 6017 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 6018 pLevel->addrLikeRep); 6019 VdbeCoverage(v); 6020 } 6021 #endif 6022 if( pLevel->iLeftJoin ){ 6023 int ws = pLoop->wsFlags; 6024 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 6025 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 6026 if( (ws & WHERE_IDX_ONLY)==0 ){ 6027 assert( pLevel->iTabCur==pTabList->a[pLevel->iFrom].iCursor ); 6028 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iTabCur); 6029 } 6030 if( (ws & WHERE_INDEXED) 6031 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCoveringIdx) 6032 ){ 6033 if( ws & WHERE_MULTI_OR ){ 6034 Index *pIx = pLevel->u.pCoveringIdx; 6035 int iDb = sqlite3SchemaToIndex(db, pIx->pSchema); 6036 sqlite3VdbeAddOp3(v, OP_ReopenIdx, pLevel->iIdxCur, pIx->tnum, iDb); 6037 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 6038 } 6039 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 6040 } 6041 if( pLevel->op==OP_Return ){ 6042 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 6043 }else{ 6044 sqlite3VdbeGoto(v, pLevel->addrFirst); 6045 } 6046 sqlite3VdbeJumpHere(v, addr); 6047 } 6048 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 6049 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 6050 } 6051 6052 /* The "break" point is here, just past the end of the outer loop. 6053 ** Set it. 6054 */ 6055 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 6056 6057 assert( pWInfo->nLevel<=pTabList->nSrc ); 6058 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 6059 int k, last; 6060 VdbeOp *pOp, *pLastOp; 6061 Index *pIdx = 0; 6062 SrcItem *pTabItem = &pTabList->a[pLevel->iFrom]; 6063 Table *pTab = pTabItem->pTab; 6064 assert( pTab!=0 ); 6065 pLoop = pLevel->pWLoop; 6066 6067 /* For a co-routine, change all OP_Column references to the table of 6068 ** the co-routine into OP_Copy of result contained in a register. 6069 ** OP_Rowid becomes OP_Null. 6070 */ 6071 if( pTabItem->fg.viaCoroutine ){ 6072 testcase( pParse->db->mallocFailed ); 6073 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 6074 pTabItem->regResult, 0); 6075 continue; 6076 } 6077 6078 #ifdef SQLITE_ENABLE_EARLY_CURSOR_CLOSE 6079 /* Close all of the cursors that were opened by sqlite3WhereBegin. 6080 ** Except, do not close cursors that will be reused by the OR optimization 6081 ** (WHERE_OR_SUBCLAUSE). And do not close the OP_OpenWrite cursors 6082 ** created for the ONEPASS optimization. 6083 */ 6084 if( (pTab->tabFlags & TF_Ephemeral)==0 6085 && !IsView(pTab) 6086 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 6087 ){ 6088 int ws = pLoop->wsFlags; 6089 if( pWInfo->eOnePass==ONEPASS_OFF && (ws & WHERE_IDX_ONLY)==0 ){ 6090 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); 6091 } 6092 if( (ws & WHERE_INDEXED)!=0 6093 && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0 6094 && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1] 6095 ){ 6096 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); 6097 } 6098 } 6099 #endif 6100 6101 /* If this scan uses an index, make VDBE code substitutions to read data 6102 ** from the index instead of from the table where possible. In some cases 6103 ** this optimization prevents the table from ever being read, which can 6104 ** yield a significant performance boost. 6105 ** 6106 ** Calls to the code generator in between sqlite3WhereBegin and 6107 ** sqlite3WhereEnd will have created code that references the table 6108 ** directly. This loop scans all that code looking for opcodes 6109 ** that reference the table and converts them into opcodes that 6110 ** reference the index. 6111 */ 6112 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 6113 pIdx = pLoop->u.btree.pIndex; 6114 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 6115 pIdx = pLevel->u.pCoveringIdx; 6116 } 6117 if( pIdx 6118 && !db->mallocFailed 6119 ){ 6120 if( pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable) ){ 6121 last = iEnd; 6122 }else{ 6123 last = pWInfo->iEndWhere; 6124 } 6125 k = pLevel->addrBody + 1; 6126 #ifdef SQLITE_DEBUG 6127 if( db->flags & SQLITE_VdbeAddopTrace ){ 6128 printf("TRANSLATE opcodes in range %d..%d\n", k, last-1); 6129 } 6130 /* Proof that the "+1" on the k value above is safe */ 6131 pOp = sqlite3VdbeGetOp(v, k - 1); 6132 assert( pOp->opcode!=OP_Column || pOp->p1!=pLevel->iTabCur ); 6133 assert( pOp->opcode!=OP_Rowid || pOp->p1!=pLevel->iTabCur ); 6134 assert( pOp->opcode!=OP_IfNullRow || pOp->p1!=pLevel->iTabCur ); 6135 #endif 6136 pOp = sqlite3VdbeGetOp(v, k); 6137 pLastOp = pOp + (last - k); 6138 assert( pOp<=pLastOp ); 6139 do{ 6140 if( pOp->p1!=pLevel->iTabCur ){ 6141 /* no-op */ 6142 }else if( pOp->opcode==OP_Column 6143 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC 6144 || pOp->opcode==OP_Offset 6145 #endif 6146 ){ 6147 int x = pOp->p2; 6148 assert( pIdx->pTable==pTab ); 6149 if( !HasRowid(pTab) ){ 6150 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 6151 x = pPk->aiColumn[x]; 6152 assert( x>=0 ); 6153 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC 6154 }else if( pOp->opcode==OP_Offset ){ 6155 /* Do not need to translate the column number */ 6156 #endif 6157 }else{ 6158 testcase( x!=sqlite3StorageColumnToTable(pTab,x) ); 6159 x = sqlite3StorageColumnToTable(pTab,x); 6160 } 6161 x = sqlite3TableColumnToIndex(pIdx, x); 6162 if( x>=0 ){ 6163 pOp->p2 = x; 6164 pOp->p1 = pLevel->iIdxCur; 6165 OpcodeRewriteTrace(db, k, pOp); 6166 } 6167 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 6168 || pWInfo->eOnePass ); 6169 }else if( pOp->opcode==OP_Rowid ){ 6170 pOp->p1 = pLevel->iIdxCur; 6171 pOp->opcode = OP_IdxRowid; 6172 OpcodeRewriteTrace(db, k, pOp); 6173 }else if( pOp->opcode==OP_IfNullRow ){ 6174 pOp->p1 = pLevel->iIdxCur; 6175 OpcodeRewriteTrace(db, k, pOp); 6176 } 6177 #ifdef SQLITE_DEBUG 6178 k++; 6179 #endif 6180 }while( (++pOp)<pLastOp ); 6181 #ifdef SQLITE_DEBUG 6182 if( db->flags & SQLITE_VdbeAddopTrace ) printf("TRANSLATE complete\n"); 6183 #endif 6184 } 6185 } 6186 6187 /* Final cleanup 6188 */ 6189 if( pWInfo->pExprMods ) whereUndoExprMods(pWInfo); 6190 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 6191 whereInfoFree(db, pWInfo); 6192 return; 6193 } 6194