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