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 21 22 /* 23 ** Trace output macros 24 */ 25 #if defined(SQLITE_TEST) || defined(SQLITE_DEBUG) 26 /***/ int sqlite3WhereTrace = 0; 27 #endif 28 #if defined(SQLITE_DEBUG) \ 29 && (defined(SQLITE_TEST) || defined(SQLITE_ENABLE_WHERETRACE)) 30 # define WHERETRACE(X) if(sqlite3WhereTrace) sqlite3DebugPrintf X 31 #else 32 # define WHERETRACE(X) 33 #endif 34 35 /* Forward reference 36 */ 37 typedef struct WhereClause WhereClause; 38 typedef struct WhereMaskSet WhereMaskSet; 39 typedef struct WhereOrInfo WhereOrInfo; 40 typedef struct WhereAndInfo WhereAndInfo; 41 typedef struct WhereCost WhereCost; 42 43 /* 44 ** The query generator uses an array of instances of this structure to 45 ** help it analyze the subexpressions of the WHERE clause. Each WHERE 46 ** clause subexpression is separated from the others by AND operators, 47 ** usually, or sometimes subexpressions separated by OR. 48 ** 49 ** All WhereTerms are collected into a single WhereClause structure. 50 ** The following identity holds: 51 ** 52 ** WhereTerm.pWC->a[WhereTerm.idx] == WhereTerm 53 ** 54 ** When a term is of the form: 55 ** 56 ** X <op> <expr> 57 ** 58 ** where X is a column name and <op> is one of certain operators, 59 ** then WhereTerm.leftCursor and WhereTerm.u.leftColumn record the 60 ** cursor number and column number for X. WhereTerm.eOperator records 61 ** the <op> using a bitmask encoding defined by WO_xxx below. The 62 ** use of a bitmask encoding for the operator allows us to search 63 ** quickly for terms that match any of several different operators. 64 ** 65 ** A WhereTerm might also be two or more subterms connected by OR: 66 ** 67 ** (t1.X <op> <expr>) OR (t1.Y <op> <expr>) OR .... 68 ** 69 ** In this second case, wtFlag as the TERM_ORINFO set and eOperator==WO_OR 70 ** and the WhereTerm.u.pOrInfo field points to auxiliary information that 71 ** is collected about the 72 ** 73 ** If a term in the WHERE clause does not match either of the two previous 74 ** categories, then eOperator==0. The WhereTerm.pExpr field is still set 75 ** to the original subexpression content and wtFlags is set up appropriately 76 ** but no other fields in the WhereTerm object are meaningful. 77 ** 78 ** When eOperator!=0, prereqRight and prereqAll record sets of cursor numbers, 79 ** but they do so indirectly. A single WhereMaskSet structure translates 80 ** cursor number into bits and the translated bit is stored in the prereq 81 ** fields. The translation is used in order to maximize the number of 82 ** bits that will fit in a Bitmask. The VDBE cursor numbers might be 83 ** spread out over the non-negative integers. For example, the cursor 84 ** numbers might be 3, 8, 9, 10, 20, 23, 41, and 45. The WhereMaskSet 85 ** translates these sparse cursor numbers into consecutive integers 86 ** beginning with 0 in order to make the best possible use of the available 87 ** bits in the Bitmask. So, in the example above, the cursor numbers 88 ** would be mapped into integers 0 through 7. 89 ** 90 ** The number of terms in a join is limited by the number of bits 91 ** in prereqRight and prereqAll. The default is 64 bits, hence SQLite 92 ** is only able to process joins with 64 or fewer tables. 93 */ 94 typedef struct WhereTerm WhereTerm; 95 struct WhereTerm { 96 Expr *pExpr; /* Pointer to the subexpression that is this term */ 97 int iParent; /* Disable pWC->a[iParent] when this term disabled */ 98 int leftCursor; /* Cursor number of X in "X <op> <expr>" */ 99 union { 100 int leftColumn; /* Column number of X in "X <op> <expr>" */ 101 WhereOrInfo *pOrInfo; /* Extra information if (eOperator & WO_OR)!=0 */ 102 WhereAndInfo *pAndInfo; /* Extra information if (eOperator& WO_AND)!=0 */ 103 } u; 104 u16 eOperator; /* A WO_xx value describing <op> */ 105 u8 wtFlags; /* TERM_xxx bit flags. See below */ 106 u8 nChild; /* Number of children that must disable us */ 107 WhereClause *pWC; /* The clause this term is part of */ 108 Bitmask prereqRight; /* Bitmask of tables used by pExpr->pRight */ 109 Bitmask prereqAll; /* Bitmask of tables referenced by pExpr */ 110 }; 111 112 /* 113 ** Allowed values of WhereTerm.wtFlags 114 */ 115 #define TERM_DYNAMIC 0x01 /* Need to call sqlite3ExprDelete(db, pExpr) */ 116 #define TERM_VIRTUAL 0x02 /* Added by the optimizer. Do not code */ 117 #define TERM_CODED 0x04 /* This term is already coded */ 118 #define TERM_COPIED 0x08 /* Has a child */ 119 #define TERM_ORINFO 0x10 /* Need to free the WhereTerm.u.pOrInfo object */ 120 #define TERM_ANDINFO 0x20 /* Need to free the WhereTerm.u.pAndInfo obj */ 121 #define TERM_OR_OK 0x40 /* Used during OR-clause processing */ 122 #ifdef SQLITE_ENABLE_STAT3 123 # define TERM_VNULL 0x80 /* Manufactured x>NULL or x<=NULL term */ 124 #else 125 # define TERM_VNULL 0x00 /* Disabled if not using stat3 */ 126 #endif 127 128 /* 129 ** An instance of the following structure holds all information about a 130 ** WHERE clause. Mostly this is a container for one or more WhereTerms. 131 ** 132 ** Explanation of pOuter: For a WHERE clause of the form 133 ** 134 ** a AND ((b AND c) OR (d AND e)) AND f 135 ** 136 ** There are separate WhereClause objects for the whole clause and for 137 ** the subclauses "(b AND c)" and "(d AND e)". The pOuter field of the 138 ** subclauses points to the WhereClause object for the whole clause. 139 */ 140 struct WhereClause { 141 Parse *pParse; /* The parser context */ 142 WhereMaskSet *pMaskSet; /* Mapping of table cursor numbers to bitmasks */ 143 WhereClause *pOuter; /* Outer conjunction */ 144 u8 op; /* Split operator. TK_AND or TK_OR */ 145 u16 wctrlFlags; /* Might include WHERE_AND_ONLY */ 146 int nTerm; /* Number of terms */ 147 int nSlot; /* Number of entries in a[] */ 148 WhereTerm *a; /* Each a[] describes a term of the WHERE cluase */ 149 #if defined(SQLITE_SMALL_STACK) 150 WhereTerm aStatic[1]; /* Initial static space for a[] */ 151 #else 152 WhereTerm aStatic[8]; /* Initial static space for a[] */ 153 #endif 154 }; 155 156 /* 157 ** A WhereTerm with eOperator==WO_OR has its u.pOrInfo pointer set to 158 ** a dynamically allocated instance of the following structure. 159 */ 160 struct WhereOrInfo { 161 WhereClause wc; /* Decomposition into subterms */ 162 Bitmask indexable; /* Bitmask of all indexable tables in the clause */ 163 }; 164 165 /* 166 ** A WhereTerm with eOperator==WO_AND has its u.pAndInfo pointer set to 167 ** a dynamically allocated instance of the following structure. 168 */ 169 struct WhereAndInfo { 170 WhereClause wc; /* The subexpression broken out */ 171 }; 172 173 /* 174 ** An instance of the following structure keeps track of a mapping 175 ** between VDBE cursor numbers and bits of the bitmasks in WhereTerm. 176 ** 177 ** The VDBE cursor numbers are small integers contained in 178 ** SrcList_item.iCursor and Expr.iTable fields. For any given WHERE 179 ** clause, the cursor numbers might not begin with 0 and they might 180 ** contain gaps in the numbering sequence. But we want to make maximum 181 ** use of the bits in our bitmasks. This structure provides a mapping 182 ** from the sparse cursor numbers into consecutive integers beginning 183 ** with 0. 184 ** 185 ** If WhereMaskSet.ix[A]==B it means that The A-th bit of a Bitmask 186 ** corresponds VDBE cursor number B. The A-th bit of a bitmask is 1<<A. 187 ** 188 ** For example, if the WHERE clause expression used these VDBE 189 ** cursors: 4, 5, 8, 29, 57, 73. Then the WhereMaskSet structure 190 ** would map those cursor numbers into bits 0 through 5. 191 ** 192 ** Note that the mapping is not necessarily ordered. In the example 193 ** above, the mapping might go like this: 4->3, 5->1, 8->2, 29->0, 194 ** 57->5, 73->4. Or one of 719 other combinations might be used. It 195 ** does not really matter. What is important is that sparse cursor 196 ** numbers all get mapped into bit numbers that begin with 0 and contain 197 ** no gaps. 198 */ 199 struct WhereMaskSet { 200 int n; /* Number of assigned cursor values */ 201 int ix[BMS]; /* Cursor assigned to each bit */ 202 }; 203 204 /* 205 ** A WhereCost object records a lookup strategy and the estimated 206 ** cost of pursuing that strategy. 207 */ 208 struct WhereCost { 209 WherePlan plan; /* The lookup strategy */ 210 double rCost; /* Overall cost of pursuing this search strategy */ 211 Bitmask used; /* Bitmask of cursors used by this plan */ 212 }; 213 214 /* 215 ** Bitmasks for the operators that indices are able to exploit. An 216 ** OR-ed combination of these values can be used when searching for 217 ** terms in the where clause. 218 */ 219 #define WO_IN 0x001 220 #define WO_EQ 0x002 221 #define WO_LT (WO_EQ<<(TK_LT-TK_EQ)) 222 #define WO_LE (WO_EQ<<(TK_LE-TK_EQ)) 223 #define WO_GT (WO_EQ<<(TK_GT-TK_EQ)) 224 #define WO_GE (WO_EQ<<(TK_GE-TK_EQ)) 225 #define WO_MATCH 0x040 226 #define WO_ISNULL 0x080 227 #define WO_OR 0x100 /* Two or more OR-connected terms */ 228 #define WO_AND 0x200 /* Two or more AND-connected terms */ 229 #define WO_EQUIV 0x400 /* Of the form A==B, both columns */ 230 #define WO_NOOP 0x800 /* This term does not restrict search space */ 231 232 #define WO_ALL 0xfff /* Mask of all possible WO_* values */ 233 #define WO_SINGLE 0x0ff /* Mask of all non-compound WO_* values */ 234 235 /* 236 ** Value for wsFlags returned by bestIndex() and stored in 237 ** WhereLevel.wsFlags. These flags determine which search 238 ** strategies are appropriate. 239 ** 240 ** The least significant 12 bits is reserved as a mask for WO_ values above. 241 ** The WhereLevel.wsFlags field is usually set to WO_IN|WO_EQ|WO_ISNULL. 242 ** But if the table is the right table of a left join, WhereLevel.wsFlags 243 ** is set to WO_IN|WO_EQ. The WhereLevel.wsFlags field can then be used as 244 ** the "op" parameter to findTerm when we are resolving equality constraints. 245 ** ISNULL constraints will then not be used on the right table of a left 246 ** join. Tickets #2177 and #2189. 247 */ 248 #define WHERE_ROWID_EQ 0x00001000 /* rowid=EXPR or rowid IN (...) */ 249 #define WHERE_ROWID_RANGE 0x00002000 /* rowid<EXPR and/or rowid>EXPR */ 250 #define WHERE_COLUMN_EQ 0x00010000 /* x=EXPR or x IN (...) or x IS NULL */ 251 #define WHERE_COLUMN_RANGE 0x00020000 /* x<EXPR and/or x>EXPR */ 252 #define WHERE_COLUMN_IN 0x00040000 /* x IN (...) */ 253 #define WHERE_COLUMN_NULL 0x00080000 /* x IS NULL */ 254 #define WHERE_INDEXED 0x000f0000 /* Anything that uses an index */ 255 #define WHERE_NOT_FULLSCAN 0x100f3000 /* Does not do a full table scan */ 256 #define WHERE_IN_ABLE 0x080f1000 /* Able to support an IN operator */ 257 #define WHERE_TOP_LIMIT 0x00100000 /* x<EXPR or x<=EXPR constraint */ 258 #define WHERE_BTM_LIMIT 0x00200000 /* x>EXPR or x>=EXPR constraint */ 259 #define WHERE_BOTH_LIMIT 0x00300000 /* Both x>EXPR and x<EXPR */ 260 #define WHERE_IDX_ONLY 0x00400000 /* Use index only - omit table */ 261 #define WHERE_ORDERED 0x00800000 /* Output will appear in correct order */ 262 #define WHERE_REVERSE 0x01000000 /* Scan in reverse order */ 263 #define WHERE_UNIQUE 0x02000000 /* Selects no more than one row */ 264 #define WHERE_ALL_UNIQUE 0x04000000 /* This and all prior have one row */ 265 #define WHERE_VIRTUALTABLE 0x08000000 /* Use virtual-table processing */ 266 #define WHERE_MULTI_OR 0x10000000 /* OR using multiple indices */ 267 #define WHERE_TEMP_INDEX 0x20000000 /* Uses an ephemeral index */ 268 #define WHERE_DISTINCT 0x40000000 /* Correct order for DISTINCT */ 269 #define WHERE_COVER_SCAN 0x80000000 /* Full scan of a covering index */ 270 271 /* 272 ** This module contains many separate subroutines that work together to 273 ** find the best indices to use for accessing a particular table in a query. 274 ** An instance of the following structure holds context information about the 275 ** index search so that it can be more easily passed between the various 276 ** routines. 277 */ 278 typedef struct WhereBestIdx WhereBestIdx; 279 struct WhereBestIdx { 280 Parse *pParse; /* Parser context */ 281 WhereClause *pWC; /* The WHERE clause */ 282 struct SrcList_item *pSrc; /* The FROM clause term to search */ 283 Bitmask notReady; /* Mask of cursors not available */ 284 Bitmask notValid; /* Cursors not available for any purpose */ 285 ExprList *pOrderBy; /* The ORDER BY clause */ 286 ExprList *pDistinct; /* The select-list if query is DISTINCT */ 287 sqlite3_index_info **ppIdxInfo; /* Index information passed to xBestIndex */ 288 int i, n; /* Which loop is being coded; # of loops */ 289 WhereLevel *aLevel; /* Info about outer loops */ 290 WhereCost cost; /* Lowest cost query plan */ 291 }; 292 293 /* 294 ** Return TRUE if the probe cost is less than the baseline cost 295 */ 296 static int compareCost(const WhereCost *pProbe, const WhereCost *pBaseline){ 297 if( pProbe->rCost<pBaseline->rCost ) return 1; 298 if( pProbe->rCost>pBaseline->rCost ) return 0; 299 if( pProbe->plan.nOBSat>pBaseline->plan.nOBSat ) return 1; 300 if( pProbe->plan.nRow<pBaseline->plan.nRow ) return 1; 301 return 0; 302 } 303 304 /* 305 ** Initialize a preallocated WhereClause structure. 306 */ 307 static void whereClauseInit( 308 WhereClause *pWC, /* The WhereClause to be initialized */ 309 Parse *pParse, /* The parsing context */ 310 WhereMaskSet *pMaskSet, /* Mapping from table cursor numbers to bitmasks */ 311 u16 wctrlFlags /* Might include WHERE_AND_ONLY */ 312 ){ 313 pWC->pParse = pParse; 314 pWC->pMaskSet = pMaskSet; 315 pWC->pOuter = 0; 316 pWC->nTerm = 0; 317 pWC->nSlot = ArraySize(pWC->aStatic); 318 pWC->a = pWC->aStatic; 319 pWC->wctrlFlags = wctrlFlags; 320 } 321 322 /* Forward reference */ 323 static void whereClauseClear(WhereClause*); 324 325 /* 326 ** Deallocate all memory associated with a WhereOrInfo object. 327 */ 328 static void whereOrInfoDelete(sqlite3 *db, WhereOrInfo *p){ 329 whereClauseClear(&p->wc); 330 sqlite3DbFree(db, p); 331 } 332 333 /* 334 ** Deallocate all memory associated with a WhereAndInfo object. 335 */ 336 static void whereAndInfoDelete(sqlite3 *db, WhereAndInfo *p){ 337 whereClauseClear(&p->wc); 338 sqlite3DbFree(db, p); 339 } 340 341 /* 342 ** Deallocate a WhereClause structure. The WhereClause structure 343 ** itself is not freed. This routine is the inverse of whereClauseInit(). 344 */ 345 static void whereClauseClear(WhereClause *pWC){ 346 int i; 347 WhereTerm *a; 348 sqlite3 *db = pWC->pParse->db; 349 for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){ 350 if( a->wtFlags & TERM_DYNAMIC ){ 351 sqlite3ExprDelete(db, a->pExpr); 352 } 353 if( a->wtFlags & TERM_ORINFO ){ 354 whereOrInfoDelete(db, a->u.pOrInfo); 355 }else if( a->wtFlags & TERM_ANDINFO ){ 356 whereAndInfoDelete(db, a->u.pAndInfo); 357 } 358 } 359 if( pWC->a!=pWC->aStatic ){ 360 sqlite3DbFree(db, pWC->a); 361 } 362 } 363 364 /* 365 ** Add a single new WhereTerm entry to the WhereClause object pWC. 366 ** The new WhereTerm object is constructed from Expr p and with wtFlags. 367 ** The index in pWC->a[] of the new WhereTerm is returned on success. 368 ** 0 is returned if the new WhereTerm could not be added due to a memory 369 ** allocation error. The memory allocation failure will be recorded in 370 ** the db->mallocFailed flag so that higher-level functions can detect it. 371 ** 372 ** This routine will increase the size of the pWC->a[] array as necessary. 373 ** 374 ** If the wtFlags argument includes TERM_DYNAMIC, then responsibility 375 ** for freeing the expression p is assumed by the WhereClause object pWC. 376 ** This is true even if this routine fails to allocate a new WhereTerm. 377 ** 378 ** WARNING: This routine might reallocate the space used to store 379 ** WhereTerms. All pointers to WhereTerms should be invalidated after 380 ** calling this routine. Such pointers may be reinitialized by referencing 381 ** the pWC->a[] array. 382 */ 383 static int whereClauseInsert(WhereClause *pWC, Expr *p, u8 wtFlags){ 384 WhereTerm *pTerm; 385 int idx; 386 testcase( wtFlags & TERM_VIRTUAL ); /* EV: R-00211-15100 */ 387 if( pWC->nTerm>=pWC->nSlot ){ 388 WhereTerm *pOld = pWC->a; 389 sqlite3 *db = pWC->pParse->db; 390 pWC->a = sqlite3DbMallocRaw(db, sizeof(pWC->a[0])*pWC->nSlot*2 ); 391 if( pWC->a==0 ){ 392 if( wtFlags & TERM_DYNAMIC ){ 393 sqlite3ExprDelete(db, p); 394 } 395 pWC->a = pOld; 396 return 0; 397 } 398 memcpy(pWC->a, pOld, sizeof(pWC->a[0])*pWC->nTerm); 399 if( pOld!=pWC->aStatic ){ 400 sqlite3DbFree(db, pOld); 401 } 402 pWC->nSlot = sqlite3DbMallocSize(db, pWC->a)/sizeof(pWC->a[0]); 403 } 404 pTerm = &pWC->a[idx = pWC->nTerm++]; 405 pTerm->pExpr = sqlite3ExprSkipCollate(p); 406 pTerm->wtFlags = wtFlags; 407 pTerm->pWC = pWC; 408 pTerm->iParent = -1; 409 return idx; 410 } 411 412 /* 413 ** This routine identifies subexpressions in the WHERE clause where 414 ** each subexpression is separated by the AND operator or some other 415 ** operator specified in the op parameter. The WhereClause structure 416 ** is filled with pointers to subexpressions. For example: 417 ** 418 ** WHERE a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22) 419 ** \________/ \_______________/ \________________/ 420 ** slot[0] slot[1] slot[2] 421 ** 422 ** The original WHERE clause in pExpr is unaltered. All this routine 423 ** does is make slot[] entries point to substructure within pExpr. 424 ** 425 ** In the previous sentence and in the diagram, "slot[]" refers to 426 ** the WhereClause.a[] array. The slot[] array grows as needed to contain 427 ** all terms of the WHERE clause. 428 */ 429 static void whereSplit(WhereClause *pWC, Expr *pExpr, int op){ 430 pWC->op = (u8)op; 431 if( pExpr==0 ) return; 432 if( pExpr->op!=op ){ 433 whereClauseInsert(pWC, pExpr, 0); 434 }else{ 435 whereSplit(pWC, pExpr->pLeft, op); 436 whereSplit(pWC, pExpr->pRight, op); 437 } 438 } 439 440 /* 441 ** Initialize an expression mask set (a WhereMaskSet object) 442 */ 443 #define initMaskSet(P) memset(P, 0, sizeof(*P)) 444 445 /* 446 ** Return the bitmask for the given cursor number. Return 0 if 447 ** iCursor is not in the set. 448 */ 449 static Bitmask getMask(WhereMaskSet *pMaskSet, int iCursor){ 450 int i; 451 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 ); 452 for(i=0; i<pMaskSet->n; i++){ 453 if( pMaskSet->ix[i]==iCursor ){ 454 return ((Bitmask)1)<<i; 455 } 456 } 457 return 0; 458 } 459 460 /* 461 ** Create a new mask for cursor iCursor. 462 ** 463 ** There is one cursor per table in the FROM clause. The number of 464 ** tables in the FROM clause is limited by a test early in the 465 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] 466 ** array will never overflow. 467 */ 468 static void createMask(WhereMaskSet *pMaskSet, int iCursor){ 469 assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); 470 pMaskSet->ix[pMaskSet->n++] = iCursor; 471 } 472 473 /* 474 ** This routine walks (recursively) an expression tree and generates 475 ** a bitmask indicating which tables are used in that expression 476 ** tree. 477 ** 478 ** In order for this routine to work, the calling function must have 479 ** previously invoked sqlite3ResolveExprNames() on the expression. See 480 ** the header comment on that routine for additional information. 481 ** The sqlite3ResolveExprNames() routines looks for column names and 482 ** sets their opcodes to TK_COLUMN and their Expr.iTable fields to 483 ** the VDBE cursor number of the table. This routine just has to 484 ** translate the cursor numbers into bitmask values and OR all 485 ** the bitmasks together. 486 */ 487 static Bitmask exprListTableUsage(WhereMaskSet*, ExprList*); 488 static Bitmask exprSelectTableUsage(WhereMaskSet*, Select*); 489 static Bitmask exprTableUsage(WhereMaskSet *pMaskSet, Expr *p){ 490 Bitmask mask = 0; 491 if( p==0 ) return 0; 492 if( p->op==TK_COLUMN ){ 493 mask = getMask(pMaskSet, p->iTable); 494 return mask; 495 } 496 mask = exprTableUsage(pMaskSet, p->pRight); 497 mask |= exprTableUsage(pMaskSet, p->pLeft); 498 if( ExprHasProperty(p, EP_xIsSelect) ){ 499 mask |= exprSelectTableUsage(pMaskSet, p->x.pSelect); 500 }else{ 501 mask |= exprListTableUsage(pMaskSet, p->x.pList); 502 } 503 return mask; 504 } 505 static Bitmask exprListTableUsage(WhereMaskSet *pMaskSet, ExprList *pList){ 506 int i; 507 Bitmask mask = 0; 508 if( pList ){ 509 for(i=0; i<pList->nExpr; i++){ 510 mask |= exprTableUsage(pMaskSet, pList->a[i].pExpr); 511 } 512 } 513 return mask; 514 } 515 static Bitmask exprSelectTableUsage(WhereMaskSet *pMaskSet, Select *pS){ 516 Bitmask mask = 0; 517 while( pS ){ 518 SrcList *pSrc = pS->pSrc; 519 mask |= exprListTableUsage(pMaskSet, pS->pEList); 520 mask |= exprListTableUsage(pMaskSet, pS->pGroupBy); 521 mask |= exprListTableUsage(pMaskSet, pS->pOrderBy); 522 mask |= exprTableUsage(pMaskSet, pS->pWhere); 523 mask |= exprTableUsage(pMaskSet, pS->pHaving); 524 if( ALWAYS(pSrc!=0) ){ 525 int i; 526 for(i=0; i<pSrc->nSrc; i++){ 527 mask |= exprSelectTableUsage(pMaskSet, pSrc->a[i].pSelect); 528 mask |= exprTableUsage(pMaskSet, pSrc->a[i].pOn); 529 } 530 } 531 pS = pS->pPrior; 532 } 533 return mask; 534 } 535 536 /* 537 ** Return TRUE if the given operator is one of the operators that is 538 ** allowed for an indexable WHERE clause term. The allowed operators are 539 ** "=", "<", ">", "<=", ">=", and "IN". 540 ** 541 ** IMPLEMENTATION-OF: R-59926-26393 To be usable by an index a term must be 542 ** of one of the following forms: column = expression column > expression 543 ** column >= expression column < expression column <= expression 544 ** expression = column expression > column expression >= column 545 ** expression < column expression <= column column IN 546 ** (expression-list) column IN (subquery) column IS NULL 547 */ 548 static int allowedOp(int op){ 549 assert( TK_GT>TK_EQ && TK_GT<TK_GE ); 550 assert( TK_LT>TK_EQ && TK_LT<TK_GE ); 551 assert( TK_LE>TK_EQ && TK_LE<TK_GE ); 552 assert( TK_GE==TK_EQ+4 ); 553 return op==TK_IN || (op>=TK_EQ && op<=TK_GE) || op==TK_ISNULL; 554 } 555 556 /* 557 ** Swap two objects of type TYPE. 558 */ 559 #define SWAP(TYPE,A,B) {TYPE t=A; A=B; B=t;} 560 561 /* 562 ** Commute a comparison operator. Expressions of the form "X op Y" 563 ** are converted into "Y op X". 564 ** 565 ** If left/right precedence rules come into play when determining the 566 ** collating 567 ** side of the comparison, it remains associated with the same side after 568 ** the commutation. So "Y collate NOCASE op X" becomes 569 ** "X op Y". This is because any collation sequence on 570 ** the left hand side of a comparison overrides any collation sequence 571 ** attached to the right. For the same reason the EP_Collate flag 572 ** is not commuted. 573 */ 574 static void exprCommute(Parse *pParse, Expr *pExpr){ 575 u16 expRight = (pExpr->pRight->flags & EP_Collate); 576 u16 expLeft = (pExpr->pLeft->flags & EP_Collate); 577 assert( allowedOp(pExpr->op) && pExpr->op!=TK_IN ); 578 if( expRight==expLeft ){ 579 /* Either X and Y both have COLLATE operator or neither do */ 580 if( expRight ){ 581 /* Both X and Y have COLLATE operators. Make sure X is always 582 ** used by clearing the EP_Collate flag from Y. */ 583 pExpr->pRight->flags &= ~EP_Collate; 584 }else if( sqlite3ExprCollSeq(pParse, pExpr->pLeft)!=0 ){ 585 /* Neither X nor Y have COLLATE operators, but X has a non-default 586 ** collating sequence. So add the EP_Collate marker on X to cause 587 ** it to be searched first. */ 588 pExpr->pLeft->flags |= EP_Collate; 589 } 590 } 591 SWAP(Expr*,pExpr->pRight,pExpr->pLeft); 592 if( pExpr->op>=TK_GT ){ 593 assert( TK_LT==TK_GT+2 ); 594 assert( TK_GE==TK_LE+2 ); 595 assert( TK_GT>TK_EQ ); 596 assert( TK_GT<TK_LE ); 597 assert( pExpr->op>=TK_GT && pExpr->op<=TK_GE ); 598 pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT; 599 } 600 } 601 602 /* 603 ** Translate from TK_xx operator to WO_xx bitmask. 604 */ 605 static u16 operatorMask(int op){ 606 u16 c; 607 assert( allowedOp(op) ); 608 if( op==TK_IN ){ 609 c = WO_IN; 610 }else if( op==TK_ISNULL ){ 611 c = WO_ISNULL; 612 }else{ 613 assert( (WO_EQ<<(op-TK_EQ)) < 0x7fff ); 614 c = (u16)(WO_EQ<<(op-TK_EQ)); 615 } 616 assert( op!=TK_ISNULL || c==WO_ISNULL ); 617 assert( op!=TK_IN || c==WO_IN ); 618 assert( op!=TK_EQ || c==WO_EQ ); 619 assert( op!=TK_LT || c==WO_LT ); 620 assert( op!=TK_LE || c==WO_LE ); 621 assert( op!=TK_GT || c==WO_GT ); 622 assert( op!=TK_GE || c==WO_GE ); 623 return c; 624 } 625 626 /* 627 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 628 ** where X is a reference to the iColumn of table iCur and <op> is one of 629 ** the WO_xx operator codes specified by the op parameter. 630 ** Return a pointer to the term. Return 0 if not found. 631 ** 632 ** The term returned might by Y=<expr> if there is another constraint in 633 ** the WHERE clause that specifies that X=Y. Any such constraints will be 634 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 635 ** aEquiv[] array holds X and all its equivalents, with each SQL variable 636 ** taking up two slots in aEquiv[]. The first slot is for the cursor number 637 ** and the second is for the column number. There are 22 slots in aEquiv[] 638 ** so that means we can look for X plus up to 10 other equivalent values. 639 ** Hence a search for X will return <expr> if X=A1 and A1=A2 and A2=A3 640 ** and ... and A9=A10 and A10=<expr>. 641 ** 642 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 643 ** then try for the one with no dependencies on <expr> - in other words where 644 ** <expr> is a constant expression of some kind. Only return entries of 645 ** the form "X <op> Y" where Y is a column in another table if no terms of 646 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 647 ** exist, try to return a term that does not use WO_EQUIV. 648 */ 649 static WhereTerm *findTerm( 650 WhereClause *pWC, /* The WHERE clause to be searched */ 651 int iCur, /* Cursor number of LHS */ 652 int iColumn, /* Column number of LHS */ 653 Bitmask notReady, /* RHS must not overlap with this mask */ 654 u32 op, /* Mask of WO_xx values describing operator */ 655 Index *pIdx /* Must be compatible with this index, if not NULL */ 656 ){ 657 WhereTerm *pTerm; /* Term being examined as possible result */ 658 WhereTerm *pResult = 0; /* The answer to return */ 659 WhereClause *pWCOrig = pWC; /* Original pWC value */ 660 int j, k; /* Loop counters */ 661 Expr *pX; /* Pointer to an expression */ 662 Parse *pParse; /* Parsing context */ 663 int iOrigCol = iColumn; /* Original value of iColumn */ 664 int nEquiv = 2; /* Number of entires in aEquiv[] */ 665 int iEquiv = 2; /* Number of entries of aEquiv[] processed so far */ 666 int aEquiv[22]; /* iCur,iColumn and up to 10 other equivalents */ 667 668 assert( iCur>=0 ); 669 aEquiv[0] = iCur; 670 aEquiv[1] = iColumn; 671 for(;;){ 672 for(pWC=pWCOrig; pWC; pWC=pWC->pOuter){ 673 for(pTerm=pWC->a, k=pWC->nTerm; k; k--, pTerm++){ 674 if( pTerm->leftCursor==iCur 675 && pTerm->u.leftColumn==iColumn 676 ){ 677 if( (pTerm->prereqRight & notReady)==0 678 && (pTerm->eOperator & op & WO_ALL)!=0 679 ){ 680 if( iOrigCol>=0 && pIdx && (pTerm->eOperator & WO_ISNULL)==0 ){ 681 CollSeq *pColl; 682 char idxaff; 683 684 pX = pTerm->pExpr; 685 pParse = pWC->pParse; 686 idxaff = pIdx->pTable->aCol[iOrigCol].affinity; 687 if( !sqlite3IndexAffinityOk(pX, idxaff) ){ 688 continue; 689 } 690 691 /* Figure out the collation sequence required from an index for 692 ** it to be useful for optimising expression pX. Store this 693 ** value in variable pColl. 694 */ 695 assert(pX->pLeft); 696 pColl = sqlite3BinaryCompareCollSeq(pParse,pX->pLeft,pX->pRight); 697 if( pColl==0 ) pColl = pParse->db->pDfltColl; 698 699 for(j=0; pIdx->aiColumn[j]!=iOrigCol; j++){ 700 if( NEVER(j>=pIdx->nColumn) ) return 0; 701 } 702 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[j]) ){ 703 continue; 704 } 705 } 706 if( pTerm->prereqRight==0 ){ 707 pResult = pTerm; 708 goto findTerm_success; 709 }else if( pResult==0 ){ 710 pResult = pTerm; 711 } 712 } 713 if( (pTerm->eOperator & WO_EQUIV)!=0 714 && nEquiv<ArraySize(aEquiv) 715 ){ 716 pX = sqlite3ExprSkipCollate(pTerm->pExpr->pRight); 717 assert( pX->op==TK_COLUMN ); 718 for(j=0; j<nEquiv; j+=2){ 719 if( aEquiv[j]==pX->iTable && aEquiv[j+1]==pX->iColumn ) break; 720 } 721 if( j==nEquiv ){ 722 aEquiv[j] = pX->iTable; 723 aEquiv[j+1] = pX->iColumn; 724 nEquiv += 2; 725 } 726 } 727 } 728 } 729 } 730 if( iEquiv>=nEquiv ) break; 731 iCur = aEquiv[iEquiv++]; 732 iColumn = aEquiv[iEquiv++]; 733 } 734 findTerm_success: 735 return pResult; 736 } 737 738 /* Forward reference */ 739 static void exprAnalyze(SrcList*, WhereClause*, int); 740 741 /* 742 ** Call exprAnalyze on all terms in a WHERE clause. 743 ** 744 ** 745 */ 746 static void exprAnalyzeAll( 747 SrcList *pTabList, /* the FROM clause */ 748 WhereClause *pWC /* the WHERE clause to be analyzed */ 749 ){ 750 int i; 751 for(i=pWC->nTerm-1; i>=0; i--){ 752 exprAnalyze(pTabList, pWC, i); 753 } 754 } 755 756 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION 757 /* 758 ** Check to see if the given expression is a LIKE or GLOB operator that 759 ** can be optimized using inequality constraints. Return TRUE if it is 760 ** so and false if not. 761 ** 762 ** In order for the operator to be optimizible, the RHS must be a string 763 ** literal that does not begin with a wildcard. 764 */ 765 static int isLikeOrGlob( 766 Parse *pParse, /* Parsing and code generating context */ 767 Expr *pExpr, /* Test this expression */ 768 Expr **ppPrefix, /* Pointer to TK_STRING expression with pattern prefix */ 769 int *pisComplete, /* True if the only wildcard is % in the last character */ 770 int *pnoCase /* True if uppercase is equivalent to lowercase */ 771 ){ 772 const char *z = 0; /* String on RHS of LIKE operator */ 773 Expr *pRight, *pLeft; /* Right and left size of LIKE operator */ 774 ExprList *pList; /* List of operands to the LIKE operator */ 775 int c; /* One character in z[] */ 776 int cnt; /* Number of non-wildcard prefix characters */ 777 char wc[3]; /* Wildcard characters */ 778 sqlite3 *db = pParse->db; /* Database connection */ 779 sqlite3_value *pVal = 0; 780 int op; /* Opcode of pRight */ 781 782 if( !sqlite3IsLikeFunction(db, pExpr, pnoCase, wc) ){ 783 return 0; 784 } 785 #ifdef SQLITE_EBCDIC 786 if( *pnoCase ) return 0; 787 #endif 788 pList = pExpr->x.pList; 789 pLeft = pList->a[1].pExpr; 790 if( pLeft->op!=TK_COLUMN 791 || sqlite3ExprAffinity(pLeft)!=SQLITE_AFF_TEXT 792 || IsVirtual(pLeft->pTab) 793 ){ 794 /* IMP: R-02065-49465 The left-hand side of the LIKE or GLOB operator must 795 ** be the name of an indexed column with TEXT affinity. */ 796 return 0; 797 } 798 assert( pLeft->iColumn!=(-1) ); /* Because IPK never has AFF_TEXT */ 799 800 pRight = pList->a[0].pExpr; 801 op = pRight->op; 802 if( op==TK_REGISTER ){ 803 op = pRight->op2; 804 } 805 if( op==TK_VARIABLE ){ 806 Vdbe *pReprepare = pParse->pReprepare; 807 int iCol = pRight->iColumn; 808 pVal = sqlite3VdbeGetValue(pReprepare, iCol, SQLITE_AFF_NONE); 809 if( pVal && sqlite3_value_type(pVal)==SQLITE_TEXT ){ 810 z = (char *)sqlite3_value_text(pVal); 811 } 812 sqlite3VdbeSetVarmask(pParse->pVdbe, iCol); 813 assert( pRight->op==TK_VARIABLE || pRight->op==TK_REGISTER ); 814 }else if( op==TK_STRING ){ 815 z = pRight->u.zToken; 816 } 817 if( z ){ 818 cnt = 0; 819 while( (c=z[cnt])!=0 && c!=wc[0] && c!=wc[1] && c!=wc[2] ){ 820 cnt++; 821 } 822 if( cnt!=0 && 255!=(u8)z[cnt-1] ){ 823 Expr *pPrefix; 824 *pisComplete = c==wc[0] && z[cnt+1]==0; 825 pPrefix = sqlite3Expr(db, TK_STRING, z); 826 if( pPrefix ) pPrefix->u.zToken[cnt] = 0; 827 *ppPrefix = pPrefix; 828 if( op==TK_VARIABLE ){ 829 Vdbe *v = pParse->pVdbe; 830 sqlite3VdbeSetVarmask(v, pRight->iColumn); 831 if( *pisComplete && pRight->u.zToken[1] ){ 832 /* If the rhs of the LIKE expression is a variable, and the current 833 ** value of the variable means there is no need to invoke the LIKE 834 ** function, then no OP_Variable will be added to the program. 835 ** This causes problems for the sqlite3_bind_parameter_name() 836 ** API. To workaround them, add a dummy OP_Variable here. 837 */ 838 int r1 = sqlite3GetTempReg(pParse); 839 sqlite3ExprCodeTarget(pParse, pRight, r1); 840 sqlite3VdbeChangeP3(v, sqlite3VdbeCurrentAddr(v)-1, 0); 841 sqlite3ReleaseTempReg(pParse, r1); 842 } 843 } 844 }else{ 845 z = 0; 846 } 847 } 848 849 sqlite3ValueFree(pVal); 850 return (z!=0); 851 } 852 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */ 853 854 855 #ifndef SQLITE_OMIT_VIRTUALTABLE 856 /* 857 ** Check to see if the given expression is of the form 858 ** 859 ** column MATCH expr 860 ** 861 ** If it is then return TRUE. If not, return FALSE. 862 */ 863 static int isMatchOfColumn( 864 Expr *pExpr /* Test this expression */ 865 ){ 866 ExprList *pList; 867 868 if( pExpr->op!=TK_FUNCTION ){ 869 return 0; 870 } 871 if( sqlite3StrICmp(pExpr->u.zToken,"match")!=0 ){ 872 return 0; 873 } 874 pList = pExpr->x.pList; 875 if( pList->nExpr!=2 ){ 876 return 0; 877 } 878 if( pList->a[1].pExpr->op != TK_COLUMN ){ 879 return 0; 880 } 881 return 1; 882 } 883 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 884 885 /* 886 ** If the pBase expression originated in the ON or USING clause of 887 ** a join, then transfer the appropriate markings over to derived. 888 */ 889 static void transferJoinMarkings(Expr *pDerived, Expr *pBase){ 890 pDerived->flags |= pBase->flags & EP_FromJoin; 891 pDerived->iRightJoinTable = pBase->iRightJoinTable; 892 } 893 894 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY) 895 /* 896 ** Analyze a term that consists of two or more OR-connected 897 ** subterms. So in: 898 ** 899 ** ... WHERE (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13) 900 ** ^^^^^^^^^^^^^^^^^^^^ 901 ** 902 ** This routine analyzes terms such as the middle term in the above example. 903 ** A WhereOrTerm object is computed and attached to the term under 904 ** analysis, regardless of the outcome of the analysis. Hence: 905 ** 906 ** WhereTerm.wtFlags |= TERM_ORINFO 907 ** WhereTerm.u.pOrInfo = a dynamically allocated WhereOrTerm object 908 ** 909 ** The term being analyzed must have two or more of OR-connected subterms. 910 ** A single subterm might be a set of AND-connected sub-subterms. 911 ** Examples of terms under analysis: 912 ** 913 ** (A) t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5 914 ** (B) x=expr1 OR expr2=x OR x=expr3 915 ** (C) t1.x=t2.y OR (t1.x=t2.z AND t1.y=15) 916 ** (D) x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*') 917 ** (E) (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6) 918 ** 919 ** CASE 1: 920 ** 921 ** If all subterms are of the form T.C=expr for some single column of C and 922 ** a single table T (as shown in example B above) then create a new virtual 923 ** term that is an equivalent IN expression. In other words, if the term 924 ** being analyzed is: 925 ** 926 ** x = expr1 OR expr2 = x OR x = expr3 927 ** 928 ** then create a new virtual term like this: 929 ** 930 ** x IN (expr1,expr2,expr3) 931 ** 932 ** CASE 2: 933 ** 934 ** If all subterms are indexable by a single table T, then set 935 ** 936 ** WhereTerm.eOperator = WO_OR 937 ** WhereTerm.u.pOrInfo->indexable |= the cursor number for table T 938 ** 939 ** A subterm is "indexable" if it is of the form 940 ** "T.C <op> <expr>" where C is any column of table T and 941 ** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN". 942 ** A subterm is also indexable if it is an AND of two or more 943 ** subsubterms at least one of which is indexable. Indexable AND 944 ** subterms have their eOperator set to WO_AND and they have 945 ** u.pAndInfo set to a dynamically allocated WhereAndTerm object. 946 ** 947 ** From another point of view, "indexable" means that the subterm could 948 ** potentially be used with an index if an appropriate index exists. 949 ** This analysis does not consider whether or not the index exists; that 950 ** is something the bestIndex() routine will determine. This analysis 951 ** only looks at whether subterms appropriate for indexing exist. 952 ** 953 ** All examples A through E above all satisfy case 2. But if a term 954 ** also statisfies case 1 (such as B) we know that the optimizer will 955 ** always prefer case 1, so in that case we pretend that case 2 is not 956 ** satisfied. 957 ** 958 ** It might be the case that multiple tables are indexable. For example, 959 ** (E) above is indexable on tables P, Q, and R. 960 ** 961 ** Terms that satisfy case 2 are candidates for lookup by using 962 ** separate indices to find rowids for each subterm and composing 963 ** the union of all rowids using a RowSet object. This is similar 964 ** to "bitmap indices" in other database engines. 965 ** 966 ** OTHERWISE: 967 ** 968 ** If neither case 1 nor case 2 apply, then leave the eOperator set to 969 ** zero. This term is not useful for search. 970 */ 971 static void exprAnalyzeOrTerm( 972 SrcList *pSrc, /* the FROM clause */ 973 WhereClause *pWC, /* the complete WHERE clause */ 974 int idxTerm /* Index of the OR-term to be analyzed */ 975 ){ 976 Parse *pParse = pWC->pParse; /* Parser context */ 977 sqlite3 *db = pParse->db; /* Database connection */ 978 WhereTerm *pTerm = &pWC->a[idxTerm]; /* The term to be analyzed */ 979 Expr *pExpr = pTerm->pExpr; /* The expression of the term */ 980 WhereMaskSet *pMaskSet = pWC->pMaskSet; /* Table use masks */ 981 int i; /* Loop counters */ 982 WhereClause *pOrWc; /* Breakup of pTerm into subterms */ 983 WhereTerm *pOrTerm; /* A Sub-term within the pOrWc */ 984 WhereOrInfo *pOrInfo; /* Additional information associated with pTerm */ 985 Bitmask chngToIN; /* Tables that might satisfy case 1 */ 986 Bitmask indexable; /* Tables that are indexable, satisfying case 2 */ 987 988 /* 989 ** Break the OR clause into its separate subterms. The subterms are 990 ** stored in a WhereClause structure containing within the WhereOrInfo 991 ** object that is attached to the original OR clause term. 992 */ 993 assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 ); 994 assert( pExpr->op==TK_OR ); 995 pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo)); 996 if( pOrInfo==0 ) return; 997 pTerm->wtFlags |= TERM_ORINFO; 998 pOrWc = &pOrInfo->wc; 999 whereClauseInit(pOrWc, pWC->pParse, pMaskSet, pWC->wctrlFlags); 1000 whereSplit(pOrWc, pExpr, TK_OR); 1001 exprAnalyzeAll(pSrc, pOrWc); 1002 if( db->mallocFailed ) return; 1003 assert( pOrWc->nTerm>=2 ); 1004 1005 /* 1006 ** Compute the set of tables that might satisfy cases 1 or 2. 1007 */ 1008 indexable = ~(Bitmask)0; 1009 chngToIN = ~(Bitmask)0; 1010 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){ 1011 if( (pOrTerm->eOperator & WO_SINGLE)==0 ){ 1012 WhereAndInfo *pAndInfo; 1013 assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 ); 1014 chngToIN = 0; 1015 pAndInfo = sqlite3DbMallocRaw(db, sizeof(*pAndInfo)); 1016 if( pAndInfo ){ 1017 WhereClause *pAndWC; 1018 WhereTerm *pAndTerm; 1019 int j; 1020 Bitmask b = 0; 1021 pOrTerm->u.pAndInfo = pAndInfo; 1022 pOrTerm->wtFlags |= TERM_ANDINFO; 1023 pOrTerm->eOperator = WO_AND; 1024 pAndWC = &pAndInfo->wc; 1025 whereClauseInit(pAndWC, pWC->pParse, pMaskSet, pWC->wctrlFlags); 1026 whereSplit(pAndWC, pOrTerm->pExpr, TK_AND); 1027 exprAnalyzeAll(pSrc, pAndWC); 1028 pAndWC->pOuter = pWC; 1029 testcase( db->mallocFailed ); 1030 if( !db->mallocFailed ){ 1031 for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){ 1032 assert( pAndTerm->pExpr ); 1033 if( allowedOp(pAndTerm->pExpr->op) ){ 1034 b |= getMask(pMaskSet, pAndTerm->leftCursor); 1035 } 1036 } 1037 } 1038 indexable &= b; 1039 } 1040 }else if( pOrTerm->wtFlags & TERM_COPIED ){ 1041 /* Skip this term for now. We revisit it when we process the 1042 ** corresponding TERM_VIRTUAL term */ 1043 }else{ 1044 Bitmask b; 1045 b = getMask(pMaskSet, pOrTerm->leftCursor); 1046 if( pOrTerm->wtFlags & TERM_VIRTUAL ){ 1047 WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent]; 1048 b |= getMask(pMaskSet, pOther->leftCursor); 1049 } 1050 indexable &= b; 1051 if( (pOrTerm->eOperator & WO_EQ)==0 ){ 1052 chngToIN = 0; 1053 }else{ 1054 chngToIN &= b; 1055 } 1056 } 1057 } 1058 1059 /* 1060 ** Record the set of tables that satisfy case 2. The set might be 1061 ** empty. 1062 */ 1063 pOrInfo->indexable = indexable; 1064 pTerm->eOperator = indexable==0 ? 0 : WO_OR; 1065 1066 /* 1067 ** chngToIN holds a set of tables that *might* satisfy case 1. But 1068 ** we have to do some additional checking to see if case 1 really 1069 ** is satisfied. 1070 ** 1071 ** chngToIN will hold either 0, 1, or 2 bits. The 0-bit case means 1072 ** that there is no possibility of transforming the OR clause into an 1073 ** IN operator because one or more terms in the OR clause contain 1074 ** something other than == on a column in the single table. The 1-bit 1075 ** case means that every term of the OR clause is of the form 1076 ** "table.column=expr" for some single table. The one bit that is set 1077 ** will correspond to the common table. We still need to check to make 1078 ** sure the same column is used on all terms. The 2-bit case is when 1079 ** the all terms are of the form "table1.column=table2.column". It 1080 ** might be possible to form an IN operator with either table1.column 1081 ** or table2.column as the LHS if either is common to every term of 1082 ** the OR clause. 1083 ** 1084 ** Note that terms of the form "table.column1=table.column2" (the 1085 ** same table on both sizes of the ==) cannot be optimized. 1086 */ 1087 if( chngToIN ){ 1088 int okToChngToIN = 0; /* True if the conversion to IN is valid */ 1089 int iColumn = -1; /* Column index on lhs of IN operator */ 1090 int iCursor = -1; /* Table cursor common to all terms */ 1091 int j = 0; /* Loop counter */ 1092 1093 /* Search for a table and column that appears on one side or the 1094 ** other of the == operator in every subterm. That table and column 1095 ** will be recorded in iCursor and iColumn. There might not be any 1096 ** such table and column. Set okToChngToIN if an appropriate table 1097 ** and column is found but leave okToChngToIN false if not found. 1098 */ 1099 for(j=0; j<2 && !okToChngToIN; j++){ 1100 pOrTerm = pOrWc->a; 1101 for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){ 1102 assert( pOrTerm->eOperator & WO_EQ ); 1103 pOrTerm->wtFlags &= ~TERM_OR_OK; 1104 if( pOrTerm->leftCursor==iCursor ){ 1105 /* This is the 2-bit case and we are on the second iteration and 1106 ** current term is from the first iteration. So skip this term. */ 1107 assert( j==1 ); 1108 continue; 1109 } 1110 if( (chngToIN & getMask(pMaskSet, pOrTerm->leftCursor))==0 ){ 1111 /* This term must be of the form t1.a==t2.b where t2 is in the 1112 ** chngToIN set but t1 is not. This term will be either preceeded 1113 ** or follwed by an inverted copy (t2.b==t1.a). Skip this term 1114 ** and use its inversion. */ 1115 testcase( pOrTerm->wtFlags & TERM_COPIED ); 1116 testcase( pOrTerm->wtFlags & TERM_VIRTUAL ); 1117 assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) ); 1118 continue; 1119 } 1120 iColumn = pOrTerm->u.leftColumn; 1121 iCursor = pOrTerm->leftCursor; 1122 break; 1123 } 1124 if( i<0 ){ 1125 /* No candidate table+column was found. This can only occur 1126 ** on the second iteration */ 1127 assert( j==1 ); 1128 assert( IsPowerOfTwo(chngToIN) ); 1129 assert( chngToIN==getMask(pMaskSet, iCursor) ); 1130 break; 1131 } 1132 testcase( j==1 ); 1133 1134 /* We have found a candidate table and column. Check to see if that 1135 ** table and column is common to every term in the OR clause */ 1136 okToChngToIN = 1; 1137 for(; i>=0 && okToChngToIN; i--, pOrTerm++){ 1138 assert( pOrTerm->eOperator & WO_EQ ); 1139 if( pOrTerm->leftCursor!=iCursor ){ 1140 pOrTerm->wtFlags &= ~TERM_OR_OK; 1141 }else if( pOrTerm->u.leftColumn!=iColumn ){ 1142 okToChngToIN = 0; 1143 }else{ 1144 int affLeft, affRight; 1145 /* If the right-hand side is also a column, then the affinities 1146 ** of both right and left sides must be such that no type 1147 ** conversions are required on the right. (Ticket #2249) 1148 */ 1149 affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight); 1150 affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft); 1151 if( affRight!=0 && affRight!=affLeft ){ 1152 okToChngToIN = 0; 1153 }else{ 1154 pOrTerm->wtFlags |= TERM_OR_OK; 1155 } 1156 } 1157 } 1158 } 1159 1160 /* At this point, okToChngToIN is true if original pTerm satisfies 1161 ** case 1. In that case, construct a new virtual term that is 1162 ** pTerm converted into an IN operator. 1163 ** 1164 ** EV: R-00211-15100 1165 */ 1166 if( okToChngToIN ){ 1167 Expr *pDup; /* A transient duplicate expression */ 1168 ExprList *pList = 0; /* The RHS of the IN operator */ 1169 Expr *pLeft = 0; /* The LHS of the IN operator */ 1170 Expr *pNew; /* The complete IN operator */ 1171 1172 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){ 1173 if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue; 1174 assert( pOrTerm->eOperator & WO_EQ ); 1175 assert( pOrTerm->leftCursor==iCursor ); 1176 assert( pOrTerm->u.leftColumn==iColumn ); 1177 pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0); 1178 pList = sqlite3ExprListAppend(pWC->pParse, pList, pDup); 1179 pLeft = pOrTerm->pExpr->pLeft; 1180 } 1181 assert( pLeft!=0 ); 1182 pDup = sqlite3ExprDup(db, pLeft, 0); 1183 pNew = sqlite3PExpr(pParse, TK_IN, pDup, 0, 0); 1184 if( pNew ){ 1185 int idxNew; 1186 transferJoinMarkings(pNew, pExpr); 1187 assert( !ExprHasProperty(pNew, EP_xIsSelect) ); 1188 pNew->x.pList = pList; 1189 idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC); 1190 testcase( idxNew==0 ); 1191 exprAnalyze(pSrc, pWC, idxNew); 1192 pTerm = &pWC->a[idxTerm]; 1193 pWC->a[idxNew].iParent = idxTerm; 1194 pTerm->nChild = 1; 1195 }else{ 1196 sqlite3ExprListDelete(db, pList); 1197 } 1198 pTerm->eOperator = WO_NOOP; /* case 1 trumps case 2 */ 1199 } 1200 } 1201 } 1202 #endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */ 1203 1204 /* 1205 ** The input to this routine is an WhereTerm structure with only the 1206 ** "pExpr" field filled in. The job of this routine is to analyze the 1207 ** subexpression and populate all the other fields of the WhereTerm 1208 ** structure. 1209 ** 1210 ** If the expression is of the form "<expr> <op> X" it gets commuted 1211 ** to the standard form of "X <op> <expr>". 1212 ** 1213 ** If the expression is of the form "X <op> Y" where both X and Y are 1214 ** columns, then the original expression is unchanged and a new virtual 1215 ** term of the form "Y <op> X" is added to the WHERE clause and 1216 ** analyzed separately. The original term is marked with TERM_COPIED 1217 ** and the new term is marked with TERM_DYNAMIC (because it's pExpr 1218 ** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it 1219 ** is a commuted copy of a prior term.) The original term has nChild=1 1220 ** and the copy has idxParent set to the index of the original term. 1221 */ 1222 static void exprAnalyze( 1223 SrcList *pSrc, /* the FROM clause */ 1224 WhereClause *pWC, /* the WHERE clause */ 1225 int idxTerm /* Index of the term to be analyzed */ 1226 ){ 1227 WhereTerm *pTerm; /* The term to be analyzed */ 1228 WhereMaskSet *pMaskSet; /* Set of table index masks */ 1229 Expr *pExpr; /* The expression to be analyzed */ 1230 Bitmask prereqLeft; /* Prerequesites of the pExpr->pLeft */ 1231 Bitmask prereqAll; /* Prerequesites of pExpr */ 1232 Bitmask extraRight = 0; /* Extra dependencies on LEFT JOIN */ 1233 Expr *pStr1 = 0; /* RHS of LIKE/GLOB operator */ 1234 int isComplete = 0; /* RHS of LIKE/GLOB ends with wildcard */ 1235 int noCase = 0; /* LIKE/GLOB distinguishes case */ 1236 int op; /* Top-level operator. pExpr->op */ 1237 Parse *pParse = pWC->pParse; /* Parsing context */ 1238 sqlite3 *db = pParse->db; /* Database connection */ 1239 1240 if( db->mallocFailed ){ 1241 return; 1242 } 1243 pTerm = &pWC->a[idxTerm]; 1244 pMaskSet = pWC->pMaskSet; 1245 pExpr = pTerm->pExpr; 1246 assert( pExpr->op!=TK_AS && pExpr->op!=TK_COLLATE ); 1247 prereqLeft = exprTableUsage(pMaskSet, pExpr->pLeft); 1248 op = pExpr->op; 1249 if( op==TK_IN ){ 1250 assert( pExpr->pRight==0 ); 1251 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 1252 pTerm->prereqRight = exprSelectTableUsage(pMaskSet, pExpr->x.pSelect); 1253 }else{ 1254 pTerm->prereqRight = exprListTableUsage(pMaskSet, pExpr->x.pList); 1255 } 1256 }else if( op==TK_ISNULL ){ 1257 pTerm->prereqRight = 0; 1258 }else{ 1259 pTerm->prereqRight = exprTableUsage(pMaskSet, pExpr->pRight); 1260 } 1261 prereqAll = exprTableUsage(pMaskSet, pExpr); 1262 if( ExprHasProperty(pExpr, EP_FromJoin) ){ 1263 Bitmask x = getMask(pMaskSet, pExpr->iRightJoinTable); 1264 prereqAll |= x; 1265 extraRight = x-1; /* ON clause terms may not be used with an index 1266 ** on left table of a LEFT JOIN. Ticket #3015 */ 1267 } 1268 pTerm->prereqAll = prereqAll; 1269 pTerm->leftCursor = -1; 1270 pTerm->iParent = -1; 1271 pTerm->eOperator = 0; 1272 if( allowedOp(op) ){ 1273 Expr *pLeft = sqlite3ExprSkipCollate(pExpr->pLeft); 1274 Expr *pRight = sqlite3ExprSkipCollate(pExpr->pRight); 1275 u16 opMask = (pTerm->prereqRight & prereqLeft)==0 ? WO_ALL : WO_EQUIV; 1276 if( pLeft->op==TK_COLUMN ){ 1277 pTerm->leftCursor = pLeft->iTable; 1278 pTerm->u.leftColumn = pLeft->iColumn; 1279 pTerm->eOperator = operatorMask(op) & opMask; 1280 } 1281 if( pRight && pRight->op==TK_COLUMN ){ 1282 WhereTerm *pNew; 1283 Expr *pDup; 1284 u16 eExtraOp = 0; /* Extra bits for pNew->eOperator */ 1285 if( pTerm->leftCursor>=0 ){ 1286 int idxNew; 1287 pDup = sqlite3ExprDup(db, pExpr, 0); 1288 if( db->mallocFailed ){ 1289 sqlite3ExprDelete(db, pDup); 1290 return; 1291 } 1292 idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC); 1293 if( idxNew==0 ) return; 1294 pNew = &pWC->a[idxNew]; 1295 pNew->iParent = idxTerm; 1296 pTerm = &pWC->a[idxTerm]; 1297 pTerm->nChild = 1; 1298 pTerm->wtFlags |= TERM_COPIED; 1299 if( pExpr->op==TK_EQ 1300 && !ExprHasProperty(pExpr, EP_FromJoin) 1301 && OptimizationEnabled(db, SQLITE_Transitive) 1302 ){ 1303 pTerm->eOperator |= WO_EQUIV; 1304 eExtraOp = WO_EQUIV; 1305 } 1306 }else{ 1307 pDup = pExpr; 1308 pNew = pTerm; 1309 } 1310 exprCommute(pParse, pDup); 1311 pLeft = sqlite3ExprSkipCollate(pDup->pLeft); 1312 pNew->leftCursor = pLeft->iTable; 1313 pNew->u.leftColumn = pLeft->iColumn; 1314 testcase( (prereqLeft | extraRight) != prereqLeft ); 1315 pNew->prereqRight = prereqLeft | extraRight; 1316 pNew->prereqAll = prereqAll; 1317 pNew->eOperator = (operatorMask(pDup->op) + eExtraOp) & opMask; 1318 } 1319 } 1320 1321 #ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION 1322 /* If a term is the BETWEEN operator, create two new virtual terms 1323 ** that define the range that the BETWEEN implements. For example: 1324 ** 1325 ** a BETWEEN b AND c 1326 ** 1327 ** is converted into: 1328 ** 1329 ** (a BETWEEN b AND c) AND (a>=b) AND (a<=c) 1330 ** 1331 ** The two new terms are added onto the end of the WhereClause object. 1332 ** The new terms are "dynamic" and are children of the original BETWEEN 1333 ** term. That means that if the BETWEEN term is coded, the children are 1334 ** skipped. Or, if the children are satisfied by an index, the original 1335 ** BETWEEN term is skipped. 1336 */ 1337 else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){ 1338 ExprList *pList = pExpr->x.pList; 1339 int i; 1340 static const u8 ops[] = {TK_GE, TK_LE}; 1341 assert( pList!=0 ); 1342 assert( pList->nExpr==2 ); 1343 for(i=0; i<2; i++){ 1344 Expr *pNewExpr; 1345 int idxNew; 1346 pNewExpr = sqlite3PExpr(pParse, ops[i], 1347 sqlite3ExprDup(db, pExpr->pLeft, 0), 1348 sqlite3ExprDup(db, pList->a[i].pExpr, 0), 0); 1349 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC); 1350 testcase( idxNew==0 ); 1351 exprAnalyze(pSrc, pWC, idxNew); 1352 pTerm = &pWC->a[idxTerm]; 1353 pWC->a[idxNew].iParent = idxTerm; 1354 } 1355 pTerm->nChild = 2; 1356 } 1357 #endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */ 1358 1359 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY) 1360 /* Analyze a term that is composed of two or more subterms connected by 1361 ** an OR operator. 1362 */ 1363 else if( pExpr->op==TK_OR ){ 1364 assert( pWC->op==TK_AND ); 1365 exprAnalyzeOrTerm(pSrc, pWC, idxTerm); 1366 pTerm = &pWC->a[idxTerm]; 1367 } 1368 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */ 1369 1370 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION 1371 /* Add constraints to reduce the search space on a LIKE or GLOB 1372 ** operator. 1373 ** 1374 ** A like pattern of the form "x LIKE 'abc%'" is changed into constraints 1375 ** 1376 ** x>='abc' AND x<'abd' AND x LIKE 'abc%' 1377 ** 1378 ** The last character of the prefix "abc" is incremented to form the 1379 ** termination condition "abd". 1380 */ 1381 if( pWC->op==TK_AND 1382 && isLikeOrGlob(pParse, pExpr, &pStr1, &isComplete, &noCase) 1383 ){ 1384 Expr *pLeft; /* LHS of LIKE/GLOB operator */ 1385 Expr *pStr2; /* Copy of pStr1 - RHS of LIKE/GLOB operator */ 1386 Expr *pNewExpr1; 1387 Expr *pNewExpr2; 1388 int idxNew1; 1389 int idxNew2; 1390 Token sCollSeqName; /* Name of collating sequence */ 1391 1392 pLeft = pExpr->x.pList->a[1].pExpr; 1393 pStr2 = sqlite3ExprDup(db, pStr1, 0); 1394 if( !db->mallocFailed ){ 1395 u8 c, *pC; /* Last character before the first wildcard */ 1396 pC = (u8*)&pStr2->u.zToken[sqlite3Strlen30(pStr2->u.zToken)-1]; 1397 c = *pC; 1398 if( noCase ){ 1399 /* The point is to increment the last character before the first 1400 ** wildcard. But if we increment '@', that will push it into the 1401 ** alphabetic range where case conversions will mess up the 1402 ** inequality. To avoid this, make sure to also run the full 1403 ** LIKE on all candidate expressions by clearing the isComplete flag 1404 */ 1405 if( c=='A'-1 ) isComplete = 0; /* EV: R-64339-08207 */ 1406 1407 1408 c = sqlite3UpperToLower[c]; 1409 } 1410 *pC = c + 1; 1411 } 1412 sCollSeqName.z = noCase ? "NOCASE" : "BINARY"; 1413 sCollSeqName.n = 6; 1414 pNewExpr1 = sqlite3ExprDup(db, pLeft, 0); 1415 pNewExpr1 = sqlite3PExpr(pParse, TK_GE, 1416 sqlite3ExprAddCollateToken(pParse,pNewExpr1,&sCollSeqName), 1417 pStr1, 0); 1418 idxNew1 = whereClauseInsert(pWC, pNewExpr1, TERM_VIRTUAL|TERM_DYNAMIC); 1419 testcase( idxNew1==0 ); 1420 exprAnalyze(pSrc, pWC, idxNew1); 1421 pNewExpr2 = sqlite3ExprDup(db, pLeft, 0); 1422 pNewExpr2 = sqlite3PExpr(pParse, TK_LT, 1423 sqlite3ExprAddCollateToken(pParse,pNewExpr2,&sCollSeqName), 1424 pStr2, 0); 1425 idxNew2 = whereClauseInsert(pWC, pNewExpr2, TERM_VIRTUAL|TERM_DYNAMIC); 1426 testcase( idxNew2==0 ); 1427 exprAnalyze(pSrc, pWC, idxNew2); 1428 pTerm = &pWC->a[idxTerm]; 1429 if( isComplete ){ 1430 pWC->a[idxNew1].iParent = idxTerm; 1431 pWC->a[idxNew2].iParent = idxTerm; 1432 pTerm->nChild = 2; 1433 } 1434 } 1435 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */ 1436 1437 #ifndef SQLITE_OMIT_VIRTUALTABLE 1438 /* Add a WO_MATCH auxiliary term to the constraint set if the 1439 ** current expression is of the form: column MATCH expr. 1440 ** This information is used by the xBestIndex methods of 1441 ** virtual tables. The native query optimizer does not attempt 1442 ** to do anything with MATCH functions. 1443 */ 1444 if( isMatchOfColumn(pExpr) ){ 1445 int idxNew; 1446 Expr *pRight, *pLeft; 1447 WhereTerm *pNewTerm; 1448 Bitmask prereqColumn, prereqExpr; 1449 1450 pRight = pExpr->x.pList->a[0].pExpr; 1451 pLeft = pExpr->x.pList->a[1].pExpr; 1452 prereqExpr = exprTableUsage(pMaskSet, pRight); 1453 prereqColumn = exprTableUsage(pMaskSet, pLeft); 1454 if( (prereqExpr & prereqColumn)==0 ){ 1455 Expr *pNewExpr; 1456 pNewExpr = sqlite3PExpr(pParse, TK_MATCH, 1457 0, sqlite3ExprDup(db, pRight, 0), 0); 1458 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC); 1459 testcase( idxNew==0 ); 1460 pNewTerm = &pWC->a[idxNew]; 1461 pNewTerm->prereqRight = prereqExpr; 1462 pNewTerm->leftCursor = pLeft->iTable; 1463 pNewTerm->u.leftColumn = pLeft->iColumn; 1464 pNewTerm->eOperator = WO_MATCH; 1465 pNewTerm->iParent = idxTerm; 1466 pTerm = &pWC->a[idxTerm]; 1467 pTerm->nChild = 1; 1468 pTerm->wtFlags |= TERM_COPIED; 1469 pNewTerm->prereqAll = pTerm->prereqAll; 1470 } 1471 } 1472 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 1473 1474 #ifdef SQLITE_ENABLE_STAT3 1475 /* When sqlite_stat3 histogram data is available an operator of the 1476 ** form "x IS NOT NULL" can sometimes be evaluated more efficiently 1477 ** as "x>NULL" if x is not an INTEGER PRIMARY KEY. So construct a 1478 ** virtual term of that form. 1479 ** 1480 ** Note that the virtual term must be tagged with TERM_VNULL. This 1481 ** TERM_VNULL tag will suppress the not-null check at the beginning 1482 ** of the loop. Without the TERM_VNULL flag, the not-null check at 1483 ** the start of the loop will prevent any results from being returned. 1484 */ 1485 if( pExpr->op==TK_NOTNULL 1486 && pExpr->pLeft->op==TK_COLUMN 1487 && pExpr->pLeft->iColumn>=0 1488 ){ 1489 Expr *pNewExpr; 1490 Expr *pLeft = pExpr->pLeft; 1491 int idxNew; 1492 WhereTerm *pNewTerm; 1493 1494 pNewExpr = sqlite3PExpr(pParse, TK_GT, 1495 sqlite3ExprDup(db, pLeft, 0), 1496 sqlite3PExpr(pParse, TK_NULL, 0, 0, 0), 0); 1497 1498 idxNew = whereClauseInsert(pWC, pNewExpr, 1499 TERM_VIRTUAL|TERM_DYNAMIC|TERM_VNULL); 1500 if( idxNew ){ 1501 pNewTerm = &pWC->a[idxNew]; 1502 pNewTerm->prereqRight = 0; 1503 pNewTerm->leftCursor = pLeft->iTable; 1504 pNewTerm->u.leftColumn = pLeft->iColumn; 1505 pNewTerm->eOperator = WO_GT; 1506 pNewTerm->iParent = idxTerm; 1507 pTerm = &pWC->a[idxTerm]; 1508 pTerm->nChild = 1; 1509 pTerm->wtFlags |= TERM_COPIED; 1510 pNewTerm->prereqAll = pTerm->prereqAll; 1511 } 1512 } 1513 #endif /* SQLITE_ENABLE_STAT */ 1514 1515 /* Prevent ON clause terms of a LEFT JOIN from being used to drive 1516 ** an index for tables to the left of the join. 1517 */ 1518 pTerm->prereqRight |= extraRight; 1519 } 1520 1521 /* 1522 ** This function searches the expression list passed as the second argument 1523 ** for an expression of type TK_COLUMN that refers to the same column and 1524 ** uses the same collation sequence as the iCol'th column of index pIdx. 1525 ** Argument iBase is the cursor number used for the table that pIdx refers 1526 ** to. 1527 ** 1528 ** If such an expression is found, its index in pList->a[] is returned. If 1529 ** no expression is found, -1 is returned. 1530 */ 1531 static int findIndexCol( 1532 Parse *pParse, /* Parse context */ 1533 ExprList *pList, /* Expression list to search */ 1534 int iBase, /* Cursor for table associated with pIdx */ 1535 Index *pIdx, /* Index to match column of */ 1536 int iCol /* Column of index to match */ 1537 ){ 1538 int i; 1539 const char *zColl = pIdx->azColl[iCol]; 1540 1541 for(i=0; i<pList->nExpr; i++){ 1542 Expr *p = sqlite3ExprSkipCollate(pList->a[i].pExpr); 1543 if( p->op==TK_COLUMN 1544 && p->iColumn==pIdx->aiColumn[iCol] 1545 && p->iTable==iBase 1546 ){ 1547 CollSeq *pColl = sqlite3ExprCollSeq(pParse, pList->a[i].pExpr); 1548 if( ALWAYS(pColl) && 0==sqlite3StrICmp(pColl->zName, zColl) ){ 1549 return i; 1550 } 1551 } 1552 } 1553 1554 return -1; 1555 } 1556 1557 /* 1558 ** This routine determines if pIdx can be used to assist in processing a 1559 ** DISTINCT qualifier. In other words, it tests whether or not using this 1560 ** index for the outer loop guarantees that rows with equal values for 1561 ** all expressions in the pDistinct list are delivered grouped together. 1562 ** 1563 ** For example, the query 1564 ** 1565 ** SELECT DISTINCT a, b, c FROM tbl WHERE a = ? 1566 ** 1567 ** can benefit from any index on columns "b" and "c". 1568 */ 1569 static int isDistinctIndex( 1570 Parse *pParse, /* Parsing context */ 1571 WhereClause *pWC, /* The WHERE clause */ 1572 Index *pIdx, /* The index being considered */ 1573 int base, /* Cursor number for the table pIdx is on */ 1574 ExprList *pDistinct, /* The DISTINCT expressions */ 1575 int nEqCol /* Number of index columns with == */ 1576 ){ 1577 Bitmask mask = 0; /* Mask of unaccounted for pDistinct exprs */ 1578 int i; /* Iterator variable */ 1579 1580 assert( pDistinct!=0 ); 1581 if( pIdx->zName==0 || pDistinct->nExpr>=BMS ) return 0; 1582 testcase( pDistinct->nExpr==BMS-1 ); 1583 1584 /* Loop through all the expressions in the distinct list. If any of them 1585 ** are not simple column references, return early. Otherwise, test if the 1586 ** WHERE clause contains a "col=X" clause. If it does, the expression 1587 ** can be ignored. If it does not, and the column does not belong to the 1588 ** same table as index pIdx, return early. Finally, if there is no 1589 ** matching "col=X" expression and the column is on the same table as pIdx, 1590 ** set the corresponding bit in variable mask. 1591 */ 1592 for(i=0; i<pDistinct->nExpr; i++){ 1593 WhereTerm *pTerm; 1594 Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr); 1595 if( p->op!=TK_COLUMN ) return 0; 1596 pTerm = findTerm(pWC, p->iTable, p->iColumn, ~(Bitmask)0, WO_EQ, 0); 1597 if( pTerm ){ 1598 Expr *pX = pTerm->pExpr; 1599 CollSeq *p1 = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); 1600 CollSeq *p2 = sqlite3ExprCollSeq(pParse, p); 1601 if( p1==p2 ) continue; 1602 } 1603 if( p->iTable!=base ) return 0; 1604 mask |= (((Bitmask)1) << i); 1605 } 1606 1607 for(i=nEqCol; mask && i<pIdx->nColumn; i++){ 1608 int iExpr = findIndexCol(pParse, pDistinct, base, pIdx, i); 1609 if( iExpr<0 ) break; 1610 mask &= ~(((Bitmask)1) << iExpr); 1611 } 1612 1613 return (mask==0); 1614 } 1615 1616 1617 /* 1618 ** Return true if the DISTINCT expression-list passed as the third argument 1619 ** is redundant. A DISTINCT list is redundant if the database contains a 1620 ** UNIQUE index that guarantees that the result of the query will be distinct 1621 ** anyway. 1622 */ 1623 static int isDistinctRedundant( 1624 Parse *pParse, 1625 SrcList *pTabList, 1626 WhereClause *pWC, 1627 ExprList *pDistinct 1628 ){ 1629 Table *pTab; 1630 Index *pIdx; 1631 int i; 1632 int iBase; 1633 1634 /* If there is more than one table or sub-select in the FROM clause of 1635 ** this query, then it will not be possible to show that the DISTINCT 1636 ** clause is redundant. */ 1637 if( pTabList->nSrc!=1 ) return 0; 1638 iBase = pTabList->a[0].iCursor; 1639 pTab = pTabList->a[0].pTab; 1640 1641 /* If any of the expressions is an IPK column on table iBase, then return 1642 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 1643 ** current SELECT is a correlated sub-query. 1644 */ 1645 for(i=0; i<pDistinct->nExpr; i++){ 1646 Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr); 1647 if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1; 1648 } 1649 1650 /* Loop through all indices on the table, checking each to see if it makes 1651 ** the DISTINCT qualifier redundant. It does so if: 1652 ** 1653 ** 1. The index is itself UNIQUE, and 1654 ** 1655 ** 2. All of the columns in the index are either part of the pDistinct 1656 ** list, or else the WHERE clause contains a term of the form "col=X", 1657 ** where X is a constant value. The collation sequences of the 1658 ** comparison and select-list expressions must match those of the index. 1659 ** 1660 ** 3. All of those index columns for which the WHERE clause does not 1661 ** contain a "col=X" term are subject to a NOT NULL constraint. 1662 */ 1663 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 1664 if( pIdx->onError==OE_None ) continue; 1665 for(i=0; i<pIdx->nColumn; i++){ 1666 int iCol = pIdx->aiColumn[i]; 1667 if( 0==findTerm(pWC, iBase, iCol, ~(Bitmask)0, WO_EQ, pIdx) ){ 1668 int iIdxCol = findIndexCol(pParse, pDistinct, iBase, pIdx, i); 1669 if( iIdxCol<0 || pTab->aCol[pIdx->aiColumn[i]].notNull==0 ){ 1670 break; 1671 } 1672 } 1673 } 1674 if( i==pIdx->nColumn ){ 1675 /* This index implies that the DISTINCT qualifier is redundant. */ 1676 return 1; 1677 } 1678 } 1679 1680 return 0; 1681 } 1682 1683 /* 1684 ** Prepare a crude estimate of the logarithm of the input value. 1685 ** The results need not be exact. This is only used for estimating 1686 ** the total cost of performing operations with O(logN) or O(NlogN) 1687 ** complexity. Because N is just a guess, it is no great tragedy if 1688 ** logN is a little off. 1689 */ 1690 static double estLog(double N){ 1691 double logN = 1; 1692 double x = 10; 1693 while( N>x ){ 1694 logN += 1; 1695 x *= 10; 1696 } 1697 return logN; 1698 } 1699 1700 /* 1701 ** Two routines for printing the content of an sqlite3_index_info 1702 ** structure. Used for testing and debugging only. If neither 1703 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 1704 ** are no-ops. 1705 */ 1706 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(SQLITE_DEBUG) 1707 static void TRACE_IDX_INPUTS(sqlite3_index_info *p){ 1708 int i; 1709 if( !sqlite3WhereTrace ) return; 1710 for(i=0; i<p->nConstraint; i++){ 1711 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", 1712 i, 1713 p->aConstraint[i].iColumn, 1714 p->aConstraint[i].iTermOffset, 1715 p->aConstraint[i].op, 1716 p->aConstraint[i].usable); 1717 } 1718 for(i=0; i<p->nOrderBy; i++){ 1719 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 1720 i, 1721 p->aOrderBy[i].iColumn, 1722 p->aOrderBy[i].desc); 1723 } 1724 } 1725 static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){ 1726 int i; 1727 if( !sqlite3WhereTrace ) return; 1728 for(i=0; i<p->nConstraint; i++){ 1729 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 1730 i, 1731 p->aConstraintUsage[i].argvIndex, 1732 p->aConstraintUsage[i].omit); 1733 } 1734 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 1735 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 1736 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 1737 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 1738 } 1739 #else 1740 #define TRACE_IDX_INPUTS(A) 1741 #define TRACE_IDX_OUTPUTS(A) 1742 #endif 1743 1744 /* 1745 ** Required because bestIndex() is called by bestOrClauseIndex() 1746 */ 1747 static void bestIndex(WhereBestIdx*); 1748 1749 /* 1750 ** This routine attempts to find an scanning strategy that can be used 1751 ** to optimize an 'OR' expression that is part of a WHERE clause. 1752 ** 1753 ** The table associated with FROM clause term pSrc may be either a 1754 ** regular B-Tree table or a virtual table. 1755 */ 1756 static void bestOrClauseIndex(WhereBestIdx *p){ 1757 #ifndef SQLITE_OMIT_OR_OPTIMIZATION 1758 WhereClause *pWC = p->pWC; /* The WHERE clause */ 1759 struct SrcList_item *pSrc = p->pSrc; /* The FROM clause term to search */ 1760 const int iCur = pSrc->iCursor; /* The cursor of the table */ 1761 const Bitmask maskSrc = getMask(pWC->pMaskSet, iCur); /* Bitmask for pSrc */ 1762 WhereTerm * const pWCEnd = &pWC->a[pWC->nTerm]; /* End of pWC->a[] */ 1763 WhereTerm *pTerm; /* A single term of the WHERE clause */ 1764 1765 /* The OR-clause optimization is disallowed if the INDEXED BY or 1766 ** NOT INDEXED clauses are used or if the WHERE_AND_ONLY bit is set. */ 1767 if( pSrc->notIndexed || pSrc->pIndex!=0 ){ 1768 return; 1769 } 1770 if( pWC->wctrlFlags & WHERE_AND_ONLY ){ 1771 return; 1772 } 1773 1774 /* Search the WHERE clause terms for a usable WO_OR term. */ 1775 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 1776 if( (pTerm->eOperator & WO_OR)!=0 1777 && ((pTerm->prereqAll & ~maskSrc) & p->notReady)==0 1778 && (pTerm->u.pOrInfo->indexable & maskSrc)!=0 1779 ){ 1780 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 1781 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 1782 WhereTerm *pOrTerm; 1783 int flags = WHERE_MULTI_OR; 1784 double rTotal = 0; 1785 double nRow = 0; 1786 Bitmask used = 0; 1787 WhereBestIdx sBOI; 1788 1789 sBOI = *p; 1790 sBOI.pOrderBy = 0; 1791 sBOI.pDistinct = 0; 1792 sBOI.ppIdxInfo = 0; 1793 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 1794 WHERETRACE(("... Multi-index OR testing for term %d of %d....\n", 1795 (pOrTerm - pOrWC->a), (pTerm - pWC->a) 1796 )); 1797 if( (pOrTerm->eOperator& WO_AND)!=0 ){ 1798 sBOI.pWC = &pOrTerm->u.pAndInfo->wc; 1799 bestIndex(&sBOI); 1800 }else if( pOrTerm->leftCursor==iCur ){ 1801 WhereClause tempWC; 1802 tempWC.pParse = pWC->pParse; 1803 tempWC.pMaskSet = pWC->pMaskSet; 1804 tempWC.pOuter = pWC; 1805 tempWC.op = TK_AND; 1806 tempWC.a = pOrTerm; 1807 tempWC.wctrlFlags = 0; 1808 tempWC.nTerm = 1; 1809 sBOI.pWC = &tempWC; 1810 bestIndex(&sBOI); 1811 }else{ 1812 continue; 1813 } 1814 rTotal += sBOI.cost.rCost; 1815 nRow += sBOI.cost.plan.nRow; 1816 used |= sBOI.cost.used; 1817 if( rTotal>=p->cost.rCost ) break; 1818 } 1819 1820 /* If there is an ORDER BY clause, increase the scan cost to account 1821 ** for the cost of the sort. */ 1822 if( p->pOrderBy!=0 ){ 1823 WHERETRACE(("... sorting increases OR cost %.9g to %.9g\n", 1824 rTotal, rTotal+nRow*estLog(nRow))); 1825 rTotal += nRow*estLog(nRow); 1826 } 1827 1828 /* If the cost of scanning using this OR term for optimization is 1829 ** less than the current cost stored in pCost, replace the contents 1830 ** of pCost. */ 1831 WHERETRACE(("... multi-index OR cost=%.9g nrow=%.9g\n", rTotal, nRow)); 1832 if( rTotal<p->cost.rCost ){ 1833 p->cost.rCost = rTotal; 1834 p->cost.used = used; 1835 p->cost.plan.nRow = nRow; 1836 p->cost.plan.nOBSat = p->i ? p->aLevel[p->i-1].plan.nOBSat : 0; 1837 p->cost.plan.wsFlags = flags; 1838 p->cost.plan.u.pTerm = pTerm; 1839 } 1840 } 1841 } 1842 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */ 1843 } 1844 1845 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 1846 /* 1847 ** Return TRUE if the WHERE clause term pTerm is of a form where it 1848 ** could be used with an index to access pSrc, assuming an appropriate 1849 ** index existed. 1850 */ 1851 static int termCanDriveIndex( 1852 WhereTerm *pTerm, /* WHERE clause term to check */ 1853 struct SrcList_item *pSrc, /* Table we are trying to access */ 1854 Bitmask notReady /* Tables in outer loops of the join */ 1855 ){ 1856 char aff; 1857 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 1858 if( (pTerm->eOperator & WO_EQ)==0 ) return 0; 1859 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 1860 aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity; 1861 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 1862 return 1; 1863 } 1864 #endif 1865 1866 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 1867 /* 1868 ** If the query plan for pSrc specified in pCost is a full table scan 1869 ** and indexing is allows (if there is no NOT INDEXED clause) and it 1870 ** possible to construct a transient index that would perform better 1871 ** than a full table scan even when the cost of constructing the index 1872 ** is taken into account, then alter the query plan to use the 1873 ** transient index. 1874 */ 1875 static void bestAutomaticIndex(WhereBestIdx *p){ 1876 Parse *pParse = p->pParse; /* The parsing context */ 1877 WhereClause *pWC = p->pWC; /* The WHERE clause */ 1878 struct SrcList_item *pSrc = p->pSrc; /* The FROM clause term to search */ 1879 double nTableRow; /* Rows in the input table */ 1880 double logN; /* log(nTableRow) */ 1881 double costTempIdx; /* per-query cost of the transient index */ 1882 WhereTerm *pTerm; /* A single term of the WHERE clause */ 1883 WhereTerm *pWCEnd; /* End of pWC->a[] */ 1884 Table *pTable; /* Table tht might be indexed */ 1885 1886 if( pParse->nQueryLoop<=(double)1 ){ 1887 /* There is no point in building an automatic index for a single scan */ 1888 return; 1889 } 1890 if( (pParse->db->flags & SQLITE_AutoIndex)==0 ){ 1891 /* Automatic indices are disabled at run-time */ 1892 return; 1893 } 1894 if( (p->cost.plan.wsFlags & WHERE_NOT_FULLSCAN)!=0 1895 && (p->cost.plan.wsFlags & WHERE_COVER_SCAN)==0 1896 ){ 1897 /* We already have some kind of index in use for this query. */ 1898 return; 1899 } 1900 if( pSrc->viaCoroutine ){ 1901 /* Cannot index a co-routine */ 1902 return; 1903 } 1904 if( pSrc->notIndexed ){ 1905 /* The NOT INDEXED clause appears in the SQL. */ 1906 return; 1907 } 1908 if( pSrc->isCorrelated ){ 1909 /* The source is a correlated sub-query. No point in indexing it. */ 1910 return; 1911 } 1912 1913 assert( pParse->nQueryLoop >= (double)1 ); 1914 pTable = pSrc->pTab; 1915 nTableRow = pTable->nRowEst; 1916 logN = estLog(nTableRow); 1917 costTempIdx = 2*logN*(nTableRow/pParse->nQueryLoop + 1); 1918 if( costTempIdx>=p->cost.rCost ){ 1919 /* The cost of creating the transient table would be greater than 1920 ** doing the full table scan */ 1921 return; 1922 } 1923 1924 /* Search for any equality comparison term */ 1925 pWCEnd = &pWC->a[pWC->nTerm]; 1926 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 1927 if( termCanDriveIndex(pTerm, pSrc, p->notReady) ){ 1928 WHERETRACE(("auto-index reduces cost from %.1f to %.1f\n", 1929 p->cost.rCost, costTempIdx)); 1930 p->cost.rCost = costTempIdx; 1931 p->cost.plan.nRow = logN + 1; 1932 p->cost.plan.wsFlags = WHERE_TEMP_INDEX; 1933 p->cost.used = pTerm->prereqRight; 1934 break; 1935 } 1936 } 1937 } 1938 #else 1939 # define bestAutomaticIndex(A) /* no-op */ 1940 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 1941 1942 1943 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 1944 /* 1945 ** Generate code to construct the Index object for an automatic index 1946 ** and to set up the WhereLevel object pLevel so that the code generator 1947 ** makes use of the automatic index. 1948 */ 1949 static void constructAutomaticIndex( 1950 Parse *pParse, /* The parsing context */ 1951 WhereClause *pWC, /* The WHERE clause */ 1952 struct SrcList_item *pSrc, /* The FROM clause term to get the next index */ 1953 Bitmask notReady, /* Mask of cursors that are not available */ 1954 WhereLevel *pLevel /* Write new index here */ 1955 ){ 1956 int nColumn; /* Number of columns in the constructed index */ 1957 WhereTerm *pTerm; /* A single term of the WHERE clause */ 1958 WhereTerm *pWCEnd; /* End of pWC->a[] */ 1959 int nByte; /* Byte of memory needed for pIdx */ 1960 Index *pIdx; /* Object describing the transient index */ 1961 Vdbe *v; /* Prepared statement under construction */ 1962 int addrInit; /* Address of the initialization bypass jump */ 1963 Table *pTable; /* The table being indexed */ 1964 KeyInfo *pKeyinfo; /* Key information for the index */ 1965 int addrTop; /* Top of the index fill loop */ 1966 int regRecord; /* Register holding an index record */ 1967 int n; /* Column counter */ 1968 int i; /* Loop counter */ 1969 int mxBitCol; /* Maximum column in pSrc->colUsed */ 1970 CollSeq *pColl; /* Collating sequence to on a column */ 1971 Bitmask idxCols; /* Bitmap of columns used for indexing */ 1972 Bitmask extraCols; /* Bitmap of additional columns */ 1973 1974 /* Generate code to skip over the creation and initialization of the 1975 ** transient index on 2nd and subsequent iterations of the loop. */ 1976 v = pParse->pVdbe; 1977 assert( v!=0 ); 1978 addrInit = sqlite3CodeOnce(pParse); 1979 1980 /* Count the number of columns that will be added to the index 1981 ** and used to match WHERE clause constraints */ 1982 nColumn = 0; 1983 pTable = pSrc->pTab; 1984 pWCEnd = &pWC->a[pWC->nTerm]; 1985 idxCols = 0; 1986 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 1987 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 1988 int iCol = pTerm->u.leftColumn; 1989 Bitmask cMask = iCol>=BMS ? ((Bitmask)1)<<(BMS-1) : ((Bitmask)1)<<iCol; 1990 testcase( iCol==BMS ); 1991 testcase( iCol==BMS-1 ); 1992 if( (idxCols & cMask)==0 ){ 1993 nColumn++; 1994 idxCols |= cMask; 1995 } 1996 } 1997 } 1998 assert( nColumn>0 ); 1999 pLevel->plan.nEq = nColumn; 2000 2001 /* Count the number of additional columns needed to create a 2002 ** covering index. A "covering index" is an index that contains all 2003 ** columns that are needed by the query. With a covering index, the 2004 ** original table never needs to be accessed. Automatic indices must 2005 ** be a covering index because the index will not be updated if the 2006 ** original table changes and the index and table cannot both be used 2007 ** if they go out of sync. 2008 */ 2009 extraCols = pSrc->colUsed & (~idxCols | (((Bitmask)1)<<(BMS-1))); 2010 mxBitCol = (pTable->nCol >= BMS-1) ? BMS-1 : pTable->nCol; 2011 testcase( pTable->nCol==BMS-1 ); 2012 testcase( pTable->nCol==BMS-2 ); 2013 for(i=0; i<mxBitCol; i++){ 2014 if( extraCols & (((Bitmask)1)<<i) ) nColumn++; 2015 } 2016 if( pSrc->colUsed & (((Bitmask)1)<<(BMS-1)) ){ 2017 nColumn += pTable->nCol - BMS + 1; 2018 } 2019 pLevel->plan.wsFlags |= WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WO_EQ; 2020 2021 /* Construct the Index object to describe this index */ 2022 nByte = sizeof(Index); 2023 nByte += nColumn*sizeof(int); /* Index.aiColumn */ 2024 nByte += nColumn*sizeof(char*); /* Index.azColl */ 2025 nByte += nColumn; /* Index.aSortOrder */ 2026 pIdx = sqlite3DbMallocZero(pParse->db, nByte); 2027 if( pIdx==0 ) return; 2028 pLevel->plan.u.pIdx = pIdx; 2029 pIdx->azColl = (char**)&pIdx[1]; 2030 pIdx->aiColumn = (int*)&pIdx->azColl[nColumn]; 2031 pIdx->aSortOrder = (u8*)&pIdx->aiColumn[nColumn]; 2032 pIdx->zName = "auto-index"; 2033 pIdx->nColumn = nColumn; 2034 pIdx->pTable = pTable; 2035 n = 0; 2036 idxCols = 0; 2037 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 2038 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 2039 int iCol = pTerm->u.leftColumn; 2040 Bitmask cMask = iCol>=BMS ? ((Bitmask)1)<<(BMS-1) : ((Bitmask)1)<<iCol; 2041 if( (idxCols & cMask)==0 ){ 2042 Expr *pX = pTerm->pExpr; 2043 idxCols |= cMask; 2044 pIdx->aiColumn[n] = pTerm->u.leftColumn; 2045 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); 2046 pIdx->azColl[n] = ALWAYS(pColl) ? pColl->zName : "BINARY"; 2047 n++; 2048 } 2049 } 2050 } 2051 assert( (u32)n==pLevel->plan.nEq ); 2052 2053 /* Add additional columns needed to make the automatic index into 2054 ** a covering index */ 2055 for(i=0; i<mxBitCol; i++){ 2056 if( extraCols & (((Bitmask)1)<<i) ){ 2057 pIdx->aiColumn[n] = i; 2058 pIdx->azColl[n] = "BINARY"; 2059 n++; 2060 } 2061 } 2062 if( pSrc->colUsed & (((Bitmask)1)<<(BMS-1)) ){ 2063 for(i=BMS-1; i<pTable->nCol; i++){ 2064 pIdx->aiColumn[n] = i; 2065 pIdx->azColl[n] = "BINARY"; 2066 n++; 2067 } 2068 } 2069 assert( n==nColumn ); 2070 2071 /* Create the automatic index */ 2072 pKeyinfo = sqlite3IndexKeyinfo(pParse, pIdx); 2073 assert( pLevel->iIdxCur>=0 ); 2074 sqlite3VdbeAddOp4(v, OP_OpenAutoindex, pLevel->iIdxCur, nColumn+1, 0, 2075 (char*)pKeyinfo, P4_KEYINFO_HANDOFF); 2076 VdbeComment((v, "for %s", pTable->zName)); 2077 2078 /* Fill the automatic index with content */ 2079 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); 2080 regRecord = sqlite3GetTempReg(pParse); 2081 sqlite3GenerateIndexKey(pParse, pIdx, pLevel->iTabCur, regRecord, 1); 2082 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 2083 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 2084 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); 2085 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 2086 sqlite3VdbeJumpHere(v, addrTop); 2087 sqlite3ReleaseTempReg(pParse, regRecord); 2088 2089 /* Jump here when skipping the initialization */ 2090 sqlite3VdbeJumpHere(v, addrInit); 2091 } 2092 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 2093 2094 #ifndef SQLITE_OMIT_VIRTUALTABLE 2095 /* 2096 ** Allocate and populate an sqlite3_index_info structure. It is the 2097 ** responsibility of the caller to eventually release the structure 2098 ** by passing the pointer returned by this function to sqlite3_free(). 2099 */ 2100 static sqlite3_index_info *allocateIndexInfo(WhereBestIdx *p){ 2101 Parse *pParse = p->pParse; 2102 WhereClause *pWC = p->pWC; 2103 struct SrcList_item *pSrc = p->pSrc; 2104 ExprList *pOrderBy = p->pOrderBy; 2105 int i, j; 2106 int nTerm; 2107 struct sqlite3_index_constraint *pIdxCons; 2108 struct sqlite3_index_orderby *pIdxOrderBy; 2109 struct sqlite3_index_constraint_usage *pUsage; 2110 WhereTerm *pTerm; 2111 int nOrderBy; 2112 sqlite3_index_info *pIdxInfo; 2113 2114 WHERETRACE(("Recomputing index info for %s...\n", pSrc->pTab->zName)); 2115 2116 /* Count the number of possible WHERE clause constraints referring 2117 ** to this virtual table */ 2118 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2119 if( pTerm->leftCursor != pSrc->iCursor ) continue; 2120 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 2121 testcase( pTerm->eOperator & WO_IN ); 2122 testcase( pTerm->eOperator & WO_ISNULL ); 2123 if( pTerm->eOperator & (WO_ISNULL) ) continue; 2124 if( pTerm->wtFlags & TERM_VNULL ) continue; 2125 nTerm++; 2126 } 2127 2128 /* If the ORDER BY clause contains only columns in the current 2129 ** virtual table then allocate space for the aOrderBy part of 2130 ** the sqlite3_index_info structure. 2131 */ 2132 nOrderBy = 0; 2133 if( pOrderBy ){ 2134 int n = pOrderBy->nExpr; 2135 for(i=0; i<n; i++){ 2136 Expr *pExpr = pOrderBy->a[i].pExpr; 2137 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; 2138 } 2139 if( i==n){ 2140 nOrderBy = n; 2141 } 2142 } 2143 2144 /* Allocate the sqlite3_index_info structure 2145 */ 2146 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 2147 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 2148 + sizeof(*pIdxOrderBy)*nOrderBy ); 2149 if( pIdxInfo==0 ){ 2150 sqlite3ErrorMsg(pParse, "out of memory"); 2151 /* (double)0 In case of SQLITE_OMIT_FLOATING_POINT... */ 2152 return 0; 2153 } 2154 2155 /* Initialize the structure. The sqlite3_index_info structure contains 2156 ** many fields that are declared "const" to prevent xBestIndex from 2157 ** changing them. We have to do some funky casting in order to 2158 ** initialize those fields. 2159 */ 2160 pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1]; 2161 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 2162 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 2163 *(int*)&pIdxInfo->nConstraint = nTerm; 2164 *(int*)&pIdxInfo->nOrderBy = nOrderBy; 2165 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons; 2166 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy; 2167 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage = 2168 pUsage; 2169 2170 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2171 u8 op; 2172 if( pTerm->leftCursor != pSrc->iCursor ) continue; 2173 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 2174 testcase( pTerm->eOperator & WO_IN ); 2175 testcase( pTerm->eOperator & WO_ISNULL ); 2176 if( pTerm->eOperator & (WO_ISNULL) ) continue; 2177 if( pTerm->wtFlags & TERM_VNULL ) continue; 2178 pIdxCons[j].iColumn = pTerm->u.leftColumn; 2179 pIdxCons[j].iTermOffset = i; 2180 op = (u8)pTerm->eOperator & WO_ALL; 2181 if( op==WO_IN ) op = WO_EQ; 2182 pIdxCons[j].op = op; 2183 /* The direct assignment in the previous line is possible only because 2184 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 2185 ** following asserts verify this fact. */ 2186 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 2187 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 2188 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 2189 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 2190 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 2191 assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH ); 2192 assert( pTerm->eOperator & (WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) ); 2193 j++; 2194 } 2195 for(i=0; i<nOrderBy; i++){ 2196 Expr *pExpr = pOrderBy->a[i].pExpr; 2197 pIdxOrderBy[i].iColumn = pExpr->iColumn; 2198 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder; 2199 } 2200 2201 return pIdxInfo; 2202 } 2203 2204 /* 2205 ** The table object reference passed as the second argument to this function 2206 ** must represent a virtual table. This function invokes the xBestIndex() 2207 ** method of the virtual table with the sqlite3_index_info pointer passed 2208 ** as the argument. 2209 ** 2210 ** If an error occurs, pParse is populated with an error message and a 2211 ** non-zero value is returned. Otherwise, 0 is returned and the output 2212 ** part of the sqlite3_index_info structure is left populated. 2213 ** 2214 ** Whether or not an error is returned, it is the responsibility of the 2215 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 2216 ** that this is required. 2217 */ 2218 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 2219 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 2220 int i; 2221 int rc; 2222 2223 WHERETRACE(("xBestIndex for %s\n", pTab->zName)); 2224 TRACE_IDX_INPUTS(p); 2225 rc = pVtab->pModule->xBestIndex(pVtab, p); 2226 TRACE_IDX_OUTPUTS(p); 2227 2228 if( rc!=SQLITE_OK ){ 2229 if( rc==SQLITE_NOMEM ){ 2230 pParse->db->mallocFailed = 1; 2231 }else if( !pVtab->zErrMsg ){ 2232 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 2233 }else{ 2234 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 2235 } 2236 } 2237 sqlite3_free(pVtab->zErrMsg); 2238 pVtab->zErrMsg = 0; 2239 2240 for(i=0; i<p->nConstraint; i++){ 2241 if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){ 2242 sqlite3ErrorMsg(pParse, 2243 "table %s: xBestIndex returned an invalid plan", pTab->zName); 2244 } 2245 } 2246 2247 return pParse->nErr; 2248 } 2249 2250 2251 /* 2252 ** Compute the best index for a virtual table. 2253 ** 2254 ** The best index is computed by the xBestIndex method of the virtual 2255 ** table module. This routine is really just a wrapper that sets up 2256 ** the sqlite3_index_info structure that is used to communicate with 2257 ** xBestIndex. 2258 ** 2259 ** In a join, this routine might be called multiple times for the 2260 ** same virtual table. The sqlite3_index_info structure is created 2261 ** and initialized on the first invocation and reused on all subsequent 2262 ** invocations. The sqlite3_index_info structure is also used when 2263 ** code is generated to access the virtual table. The whereInfoDelete() 2264 ** routine takes care of freeing the sqlite3_index_info structure after 2265 ** everybody has finished with it. 2266 */ 2267 static void bestVirtualIndex(WhereBestIdx *p){ 2268 Parse *pParse = p->pParse; /* The parsing context */ 2269 WhereClause *pWC = p->pWC; /* The WHERE clause */ 2270 struct SrcList_item *pSrc = p->pSrc; /* The FROM clause term to search */ 2271 Table *pTab = pSrc->pTab; 2272 sqlite3_index_info *pIdxInfo; 2273 struct sqlite3_index_constraint *pIdxCons; 2274 struct sqlite3_index_constraint_usage *pUsage; 2275 WhereTerm *pTerm; 2276 int i, j, k; 2277 int nOrderBy; 2278 int sortOrder; /* Sort order for IN clauses */ 2279 int bAllowIN; /* Allow IN optimizations */ 2280 double rCost; 2281 2282 /* Make sure wsFlags is initialized to some sane value. Otherwise, if the 2283 ** malloc in allocateIndexInfo() fails and this function returns leaving 2284 ** wsFlags in an uninitialized state, the caller may behave unpredictably. 2285 */ 2286 memset(&p->cost, 0, sizeof(p->cost)); 2287 p->cost.plan.wsFlags = WHERE_VIRTUALTABLE; 2288 2289 /* If the sqlite3_index_info structure has not been previously 2290 ** allocated and initialized, then allocate and initialize it now. 2291 */ 2292 pIdxInfo = *p->ppIdxInfo; 2293 if( pIdxInfo==0 ){ 2294 *p->ppIdxInfo = pIdxInfo = allocateIndexInfo(p); 2295 } 2296 if( pIdxInfo==0 ){ 2297 return; 2298 } 2299 2300 /* At this point, the sqlite3_index_info structure that pIdxInfo points 2301 ** to will have been initialized, either during the current invocation or 2302 ** during some prior invocation. Now we just have to customize the 2303 ** details of pIdxInfo for the current invocation and pass it to 2304 ** xBestIndex. 2305 */ 2306 2307 /* The module name must be defined. Also, by this point there must 2308 ** be a pointer to an sqlite3_vtab structure. Otherwise 2309 ** sqlite3ViewGetColumnNames() would have picked up the error. 2310 */ 2311 assert( pTab->azModuleArg && pTab->azModuleArg[0] ); 2312 assert( sqlite3GetVTable(pParse->db, pTab) ); 2313 2314 /* Try once or twice. On the first attempt, allow IN optimizations. 2315 ** If an IN optimization is accepted by the virtual table xBestIndex 2316 ** method, but the pInfo->aConstrainUsage.omit flag is not set, then 2317 ** the query will not work because it might allow duplicate rows in 2318 ** output. In that case, run the xBestIndex method a second time 2319 ** without the IN constraints. Usually this loop only runs once. 2320 ** The loop will exit using a "break" statement. 2321 */ 2322 for(bAllowIN=1; 1; bAllowIN--){ 2323 assert( bAllowIN==0 || bAllowIN==1 ); 2324 2325 /* Set the aConstraint[].usable fields and initialize all 2326 ** output variables to zero. 2327 ** 2328 ** aConstraint[].usable is true for constraints where the right-hand 2329 ** side contains only references to tables to the left of the current 2330 ** table. In other words, if the constraint is of the form: 2331 ** 2332 ** column = expr 2333 ** 2334 ** and we are evaluating a join, then the constraint on column is 2335 ** only valid if all tables referenced in expr occur to the left 2336 ** of the table containing column. 2337 ** 2338 ** The aConstraints[] array contains entries for all constraints 2339 ** on the current table. That way we only have to compute it once 2340 ** even though we might try to pick the best index multiple times. 2341 ** For each attempt at picking an index, the order of tables in the 2342 ** join might be different so we have to recompute the usable flag 2343 ** each time. 2344 */ 2345 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 2346 pUsage = pIdxInfo->aConstraintUsage; 2347 for(i=0; i<pIdxInfo->nConstraint; i++, pIdxCons++){ 2348 j = pIdxCons->iTermOffset; 2349 pTerm = &pWC->a[j]; 2350 if( (pTerm->prereqRight&p->notReady)==0 2351 && (bAllowIN || (pTerm->eOperator & WO_IN)==0) 2352 ){ 2353 pIdxCons->usable = 1; 2354 }else{ 2355 pIdxCons->usable = 0; 2356 } 2357 } 2358 memset(pUsage, 0, sizeof(pUsage[0])*pIdxInfo->nConstraint); 2359 if( pIdxInfo->needToFreeIdxStr ){ 2360 sqlite3_free(pIdxInfo->idxStr); 2361 } 2362 pIdxInfo->idxStr = 0; 2363 pIdxInfo->idxNum = 0; 2364 pIdxInfo->needToFreeIdxStr = 0; 2365 pIdxInfo->orderByConsumed = 0; 2366 /* ((double)2) In case of SQLITE_OMIT_FLOATING_POINT... */ 2367 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / ((double)2); 2368 nOrderBy = pIdxInfo->nOrderBy; 2369 if( !p->pOrderBy ){ 2370 pIdxInfo->nOrderBy = 0; 2371 } 2372 2373 if( vtabBestIndex(pParse, pTab, pIdxInfo) ){ 2374 return; 2375 } 2376 2377 sortOrder = SQLITE_SO_ASC; 2378 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 2379 for(i=0; i<pIdxInfo->nConstraint; i++, pIdxCons++){ 2380 if( pUsage[i].argvIndex>0 ){ 2381 j = pIdxCons->iTermOffset; 2382 pTerm = &pWC->a[j]; 2383 p->cost.used |= pTerm->prereqRight; 2384 if( (pTerm->eOperator & WO_IN)!=0 ){ 2385 if( pUsage[i].omit==0 ){ 2386 /* Do not attempt to use an IN constraint if the virtual table 2387 ** says that the equivalent EQ constraint cannot be safely omitted. 2388 ** If we do attempt to use such a constraint, some rows might be 2389 ** repeated in the output. */ 2390 break; 2391 } 2392 for(k=0; k<pIdxInfo->nOrderBy; k++){ 2393 if( pIdxInfo->aOrderBy[k].iColumn==pIdxCons->iColumn ){ 2394 sortOrder = pIdxInfo->aOrderBy[k].desc; 2395 break; 2396 } 2397 } 2398 } 2399 } 2400 } 2401 if( i>=pIdxInfo->nConstraint ) break; 2402 } 2403 2404 /* If there is an ORDER BY clause, and the selected virtual table index 2405 ** does not satisfy it, increase the cost of the scan accordingly. This 2406 ** matches the processing for non-virtual tables in bestBtreeIndex(). 2407 */ 2408 rCost = pIdxInfo->estimatedCost; 2409 if( p->pOrderBy && pIdxInfo->orderByConsumed==0 ){ 2410 rCost += estLog(rCost)*rCost; 2411 } 2412 2413 /* The cost is not allowed to be larger than SQLITE_BIG_DBL (the 2414 ** inital value of lowestCost in this loop. If it is, then the 2415 ** (cost<lowestCost) test below will never be true. 2416 ** 2417 ** Use "(double)2" instead of "2.0" in case OMIT_FLOATING_POINT 2418 ** is defined. 2419 */ 2420 if( (SQLITE_BIG_DBL/((double)2))<rCost ){ 2421 p->cost.rCost = (SQLITE_BIG_DBL/((double)2)); 2422 }else{ 2423 p->cost.rCost = rCost; 2424 } 2425 p->cost.plan.u.pVtabIdx = pIdxInfo; 2426 if( pIdxInfo->orderByConsumed ){ 2427 assert( sortOrder==0 || sortOrder==1 ); 2428 p->cost.plan.wsFlags |= WHERE_ORDERED + sortOrder*WHERE_REVERSE; 2429 p->cost.plan.nOBSat = nOrderBy; 2430 }else{ 2431 p->cost.plan.nOBSat = p->i ? p->aLevel[p->i-1].plan.nOBSat : 0; 2432 } 2433 p->cost.plan.nEq = 0; 2434 pIdxInfo->nOrderBy = nOrderBy; 2435 2436 /* Try to find a more efficient access pattern by using multiple indexes 2437 ** to optimize an OR expression within the WHERE clause. 2438 */ 2439 bestOrClauseIndex(p); 2440 } 2441 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 2442 2443 #ifdef SQLITE_ENABLE_STAT3 2444 /* 2445 ** Estimate the location of a particular key among all keys in an 2446 ** index. Store the results in aStat as follows: 2447 ** 2448 ** aStat[0] Est. number of rows less than pVal 2449 ** aStat[1] Est. number of rows equal to pVal 2450 ** 2451 ** Return SQLITE_OK on success. 2452 */ 2453 static int whereKeyStats( 2454 Parse *pParse, /* Database connection */ 2455 Index *pIdx, /* Index to consider domain of */ 2456 sqlite3_value *pVal, /* Value to consider */ 2457 int roundUp, /* Round up if true. Round down if false */ 2458 tRowcnt *aStat /* OUT: stats written here */ 2459 ){ 2460 tRowcnt n; 2461 IndexSample *aSample; 2462 int i, eType; 2463 int isEq = 0; 2464 i64 v; 2465 double r, rS; 2466 2467 assert( roundUp==0 || roundUp==1 ); 2468 assert( pIdx->nSample>0 ); 2469 if( pVal==0 ) return SQLITE_ERROR; 2470 n = pIdx->aiRowEst[0]; 2471 aSample = pIdx->aSample; 2472 eType = sqlite3_value_type(pVal); 2473 2474 if( eType==SQLITE_INTEGER ){ 2475 v = sqlite3_value_int64(pVal); 2476 r = (i64)v; 2477 for(i=0; i<pIdx->nSample; i++){ 2478 if( aSample[i].eType==SQLITE_NULL ) continue; 2479 if( aSample[i].eType>=SQLITE_TEXT ) break; 2480 if( aSample[i].eType==SQLITE_INTEGER ){ 2481 if( aSample[i].u.i>=v ){ 2482 isEq = aSample[i].u.i==v; 2483 break; 2484 } 2485 }else{ 2486 assert( aSample[i].eType==SQLITE_FLOAT ); 2487 if( aSample[i].u.r>=r ){ 2488 isEq = aSample[i].u.r==r; 2489 break; 2490 } 2491 } 2492 } 2493 }else if( eType==SQLITE_FLOAT ){ 2494 r = sqlite3_value_double(pVal); 2495 for(i=0; i<pIdx->nSample; i++){ 2496 if( aSample[i].eType==SQLITE_NULL ) continue; 2497 if( aSample[i].eType>=SQLITE_TEXT ) break; 2498 if( aSample[i].eType==SQLITE_FLOAT ){ 2499 rS = aSample[i].u.r; 2500 }else{ 2501 rS = aSample[i].u.i; 2502 } 2503 if( rS>=r ){ 2504 isEq = rS==r; 2505 break; 2506 } 2507 } 2508 }else if( eType==SQLITE_NULL ){ 2509 i = 0; 2510 if( aSample[0].eType==SQLITE_NULL ) isEq = 1; 2511 }else{ 2512 assert( eType==SQLITE_TEXT || eType==SQLITE_BLOB ); 2513 for(i=0; i<pIdx->nSample; i++){ 2514 if( aSample[i].eType==SQLITE_TEXT || aSample[i].eType==SQLITE_BLOB ){ 2515 break; 2516 } 2517 } 2518 if( i<pIdx->nSample ){ 2519 sqlite3 *db = pParse->db; 2520 CollSeq *pColl; 2521 const u8 *z; 2522 if( eType==SQLITE_BLOB ){ 2523 z = (const u8 *)sqlite3_value_blob(pVal); 2524 pColl = db->pDfltColl; 2525 assert( pColl->enc==SQLITE_UTF8 ); 2526 }else{ 2527 pColl = sqlite3GetCollSeq(pParse, SQLITE_UTF8, 0, *pIdx->azColl); 2528 if( pColl==0 ){ 2529 return SQLITE_ERROR; 2530 } 2531 z = (const u8 *)sqlite3ValueText(pVal, pColl->enc); 2532 if( !z ){ 2533 return SQLITE_NOMEM; 2534 } 2535 assert( z && pColl && pColl->xCmp ); 2536 } 2537 n = sqlite3ValueBytes(pVal, pColl->enc); 2538 2539 for(; i<pIdx->nSample; i++){ 2540 int c; 2541 int eSampletype = aSample[i].eType; 2542 if( eSampletype<eType ) continue; 2543 if( eSampletype!=eType ) break; 2544 #ifndef SQLITE_OMIT_UTF16 2545 if( pColl->enc!=SQLITE_UTF8 ){ 2546 int nSample; 2547 char *zSample = sqlite3Utf8to16( 2548 db, pColl->enc, aSample[i].u.z, aSample[i].nByte, &nSample 2549 ); 2550 if( !zSample ){ 2551 assert( db->mallocFailed ); 2552 return SQLITE_NOMEM; 2553 } 2554 c = pColl->xCmp(pColl->pUser, nSample, zSample, n, z); 2555 sqlite3DbFree(db, zSample); 2556 }else 2557 #endif 2558 { 2559 c = pColl->xCmp(pColl->pUser, aSample[i].nByte, aSample[i].u.z, n, z); 2560 } 2561 if( c>=0 ){ 2562 if( c==0 ) isEq = 1; 2563 break; 2564 } 2565 } 2566 } 2567 } 2568 2569 /* At this point, aSample[i] is the first sample that is greater than 2570 ** or equal to pVal. Or if i==pIdx->nSample, then all samples are less 2571 ** than pVal. If aSample[i]==pVal, then isEq==1. 2572 */ 2573 if( isEq ){ 2574 assert( i<pIdx->nSample ); 2575 aStat[0] = aSample[i].nLt; 2576 aStat[1] = aSample[i].nEq; 2577 }else{ 2578 tRowcnt iLower, iUpper, iGap; 2579 if( i==0 ){ 2580 iLower = 0; 2581 iUpper = aSample[0].nLt; 2582 }else{ 2583 iUpper = i>=pIdx->nSample ? n : aSample[i].nLt; 2584 iLower = aSample[i-1].nEq + aSample[i-1].nLt; 2585 } 2586 aStat[1] = pIdx->avgEq; 2587 if( iLower>=iUpper ){ 2588 iGap = 0; 2589 }else{ 2590 iGap = iUpper - iLower; 2591 } 2592 if( roundUp ){ 2593 iGap = (iGap*2)/3; 2594 }else{ 2595 iGap = iGap/3; 2596 } 2597 aStat[0] = iLower + iGap; 2598 } 2599 return SQLITE_OK; 2600 } 2601 #endif /* SQLITE_ENABLE_STAT3 */ 2602 2603 /* 2604 ** If expression pExpr represents a literal value, set *pp to point to 2605 ** an sqlite3_value structure containing the same value, with affinity 2606 ** aff applied to it, before returning. It is the responsibility of the 2607 ** caller to eventually release this structure by passing it to 2608 ** sqlite3ValueFree(). 2609 ** 2610 ** If the current parse is a recompile (sqlite3Reprepare()) and pExpr 2611 ** is an SQL variable that currently has a non-NULL value bound to it, 2612 ** create an sqlite3_value structure containing this value, again with 2613 ** affinity aff applied to it, instead. 2614 ** 2615 ** If neither of the above apply, set *pp to NULL. 2616 ** 2617 ** If an error occurs, return an error code. Otherwise, SQLITE_OK. 2618 */ 2619 #ifdef SQLITE_ENABLE_STAT3 2620 static int valueFromExpr( 2621 Parse *pParse, 2622 Expr *pExpr, 2623 u8 aff, 2624 sqlite3_value **pp 2625 ){ 2626 if( pExpr->op==TK_VARIABLE 2627 || (pExpr->op==TK_REGISTER && pExpr->op2==TK_VARIABLE) 2628 ){ 2629 int iVar = pExpr->iColumn; 2630 sqlite3VdbeSetVarmask(pParse->pVdbe, iVar); 2631 *pp = sqlite3VdbeGetValue(pParse->pReprepare, iVar, aff); 2632 return SQLITE_OK; 2633 } 2634 return sqlite3ValueFromExpr(pParse->db, pExpr, SQLITE_UTF8, aff, pp); 2635 } 2636 #endif 2637 2638 /* 2639 ** This function is used to estimate the number of rows that will be visited 2640 ** by scanning an index for a range of values. The range may have an upper 2641 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 2642 ** and lower bounds are represented by pLower and pUpper respectively. For 2643 ** example, assuming that index p is on t1(a): 2644 ** 2645 ** ... FROM t1 WHERE a > ? AND a < ? ... 2646 ** |_____| |_____| 2647 ** | | 2648 ** pLower pUpper 2649 ** 2650 ** If either of the upper or lower bound is not present, then NULL is passed in 2651 ** place of the corresponding WhereTerm. 2652 ** 2653 ** The nEq parameter is passed the index of the index column subject to the 2654 ** range constraint. Or, equivalently, the number of equality constraints 2655 ** optimized by the proposed index scan. For example, assuming index p is 2656 ** on t1(a, b), and the SQL query is: 2657 ** 2658 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 2659 ** 2660 ** then nEq should be passed the value 1 (as the range restricted column, 2661 ** b, is the second left-most column of the index). Or, if the query is: 2662 ** 2663 ** ... FROM t1 WHERE a > ? AND a < ? ... 2664 ** 2665 ** then nEq should be passed 0. 2666 ** 2667 ** The returned value is an integer divisor to reduce the estimated 2668 ** search space. A return value of 1 means that range constraints are 2669 ** no help at all. A return value of 2 means range constraints are 2670 ** expected to reduce the search space by half. And so forth... 2671 ** 2672 ** In the absence of sqlite_stat3 ANALYZE data, each range inequality 2673 ** reduces the search space by a factor of 4. Hence a single constraint (x>?) 2674 ** results in a return of 4 and a range constraint (x>? AND x<?) results 2675 ** in a return of 16. 2676 */ 2677 static int whereRangeScanEst( 2678 Parse *pParse, /* Parsing & code generating context */ 2679 Index *p, /* The index containing the range-compared column; "x" */ 2680 int nEq, /* index into p->aCol[] of the range-compared column */ 2681 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 2682 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 2683 double *pRangeDiv /* OUT: Reduce search space by this divisor */ 2684 ){ 2685 int rc = SQLITE_OK; 2686 2687 #ifdef SQLITE_ENABLE_STAT3 2688 2689 if( nEq==0 && p->nSample ){ 2690 sqlite3_value *pRangeVal; 2691 tRowcnt iLower = 0; 2692 tRowcnt iUpper = p->aiRowEst[0]; 2693 tRowcnt a[2]; 2694 u8 aff = p->pTable->aCol[p->aiColumn[0]].affinity; 2695 2696 if( pLower ){ 2697 Expr *pExpr = pLower->pExpr->pRight; 2698 rc = valueFromExpr(pParse, pExpr, aff, &pRangeVal); 2699 assert( (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 2700 if( rc==SQLITE_OK 2701 && whereKeyStats(pParse, p, pRangeVal, 0, a)==SQLITE_OK 2702 ){ 2703 iLower = a[0]; 2704 if( (pLower->eOperator & WO_GT)!=0 ) iLower += a[1]; 2705 } 2706 sqlite3ValueFree(pRangeVal); 2707 } 2708 if( rc==SQLITE_OK && pUpper ){ 2709 Expr *pExpr = pUpper->pExpr->pRight; 2710 rc = valueFromExpr(pParse, pExpr, aff, &pRangeVal); 2711 assert( (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 2712 if( rc==SQLITE_OK 2713 && whereKeyStats(pParse, p, pRangeVal, 1, a)==SQLITE_OK 2714 ){ 2715 iUpper = a[0]; 2716 if( (pUpper->eOperator & WO_LE)!=0 ) iUpper += a[1]; 2717 } 2718 sqlite3ValueFree(pRangeVal); 2719 } 2720 if( rc==SQLITE_OK ){ 2721 if( iUpper<=iLower ){ 2722 *pRangeDiv = (double)p->aiRowEst[0]; 2723 }else{ 2724 *pRangeDiv = (double)p->aiRowEst[0]/(double)(iUpper - iLower); 2725 } 2726 WHERETRACE(("range scan regions: %u..%u div=%g\n", 2727 (u32)iLower, (u32)iUpper, *pRangeDiv)); 2728 return SQLITE_OK; 2729 } 2730 } 2731 #else 2732 UNUSED_PARAMETER(pParse); 2733 UNUSED_PARAMETER(p); 2734 UNUSED_PARAMETER(nEq); 2735 #endif 2736 assert( pLower || pUpper ); 2737 *pRangeDiv = (double)1; 2738 if( pLower && (pLower->wtFlags & TERM_VNULL)==0 ) *pRangeDiv *= (double)4; 2739 if( pUpper ) *pRangeDiv *= (double)4; 2740 return rc; 2741 } 2742 2743 #ifdef SQLITE_ENABLE_STAT3 2744 /* 2745 ** Estimate the number of rows that will be returned based on 2746 ** an equality constraint x=VALUE and where that VALUE occurs in 2747 ** the histogram data. This only works when x is the left-most 2748 ** column of an index and sqlite_stat3 histogram data is available 2749 ** for that index. When pExpr==NULL that means the constraint is 2750 ** "x IS NULL" instead of "x=VALUE". 2751 ** 2752 ** Write the estimated row count into *pnRow and return SQLITE_OK. 2753 ** If unable to make an estimate, leave *pnRow unchanged and return 2754 ** non-zero. 2755 ** 2756 ** This routine can fail if it is unable to load a collating sequence 2757 ** required for string comparison, or if unable to allocate memory 2758 ** for a UTF conversion required for comparison. The error is stored 2759 ** in the pParse structure. 2760 */ 2761 static int whereEqualScanEst( 2762 Parse *pParse, /* Parsing & code generating context */ 2763 Index *p, /* The index whose left-most column is pTerm */ 2764 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 2765 double *pnRow /* Write the revised row estimate here */ 2766 ){ 2767 sqlite3_value *pRhs = 0; /* VALUE on right-hand side of pTerm */ 2768 u8 aff; /* Column affinity */ 2769 int rc; /* Subfunction return code */ 2770 tRowcnt a[2]; /* Statistics */ 2771 2772 assert( p->aSample!=0 ); 2773 assert( p->nSample>0 ); 2774 aff = p->pTable->aCol[p->aiColumn[0]].affinity; 2775 if( pExpr ){ 2776 rc = valueFromExpr(pParse, pExpr, aff, &pRhs); 2777 if( rc ) goto whereEqualScanEst_cancel; 2778 }else{ 2779 pRhs = sqlite3ValueNew(pParse->db); 2780 } 2781 if( pRhs==0 ) return SQLITE_NOTFOUND; 2782 rc = whereKeyStats(pParse, p, pRhs, 0, a); 2783 if( rc==SQLITE_OK ){ 2784 WHERETRACE(("equality scan regions: %d\n", (int)a[1])); 2785 *pnRow = a[1]; 2786 } 2787 whereEqualScanEst_cancel: 2788 sqlite3ValueFree(pRhs); 2789 return rc; 2790 } 2791 #endif /* defined(SQLITE_ENABLE_STAT3) */ 2792 2793 #ifdef SQLITE_ENABLE_STAT3 2794 /* 2795 ** Estimate the number of rows that will be returned based on 2796 ** an IN constraint where the right-hand side of the IN operator 2797 ** is a list of values. Example: 2798 ** 2799 ** WHERE x IN (1,2,3,4) 2800 ** 2801 ** Write the estimated row count into *pnRow and return SQLITE_OK. 2802 ** If unable to make an estimate, leave *pnRow unchanged and return 2803 ** non-zero. 2804 ** 2805 ** This routine can fail if it is unable to load a collating sequence 2806 ** required for string comparison, or if unable to allocate memory 2807 ** for a UTF conversion required for comparison. The error is stored 2808 ** in the pParse structure. 2809 */ 2810 static int whereInScanEst( 2811 Parse *pParse, /* Parsing & code generating context */ 2812 Index *p, /* The index whose left-most column is pTerm */ 2813 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 2814 double *pnRow /* Write the revised row estimate here */ 2815 ){ 2816 int rc = SQLITE_OK; /* Subfunction return code */ 2817 double nEst; /* Number of rows for a single term */ 2818 double nRowEst = (double)0; /* New estimate of the number of rows */ 2819 int i; /* Loop counter */ 2820 2821 assert( p->aSample!=0 ); 2822 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 2823 nEst = p->aiRowEst[0]; 2824 rc = whereEqualScanEst(pParse, p, pList->a[i].pExpr, &nEst); 2825 nRowEst += nEst; 2826 } 2827 if( rc==SQLITE_OK ){ 2828 if( nRowEst > p->aiRowEst[0] ) nRowEst = p->aiRowEst[0]; 2829 *pnRow = nRowEst; 2830 WHERETRACE(("IN row estimate: est=%g\n", nRowEst)); 2831 } 2832 return rc; 2833 } 2834 #endif /* defined(SQLITE_ENABLE_STAT3) */ 2835 2836 /* 2837 ** Check to see if column iCol of the table with cursor iTab will appear 2838 ** in sorted order according to the current query plan. 2839 ** 2840 ** Return values: 2841 ** 2842 ** 0 iCol is not ordered 2843 ** 1 iCol has only a single value 2844 ** 2 iCol is in ASC order 2845 ** 3 iCol is in DESC order 2846 */ 2847 static int isOrderedColumn( 2848 WhereBestIdx *p, 2849 int iTab, 2850 int iCol 2851 ){ 2852 int i, j; 2853 WhereLevel *pLevel = &p->aLevel[p->i-1]; 2854 Index *pIdx; 2855 u8 sortOrder; 2856 for(i=p->i-1; i>=0; i--, pLevel--){ 2857 if( pLevel->iTabCur!=iTab ) continue; 2858 if( (pLevel->plan.wsFlags & WHERE_ALL_UNIQUE)!=0 ){ 2859 return 1; 2860 } 2861 assert( (pLevel->plan.wsFlags & WHERE_ORDERED)!=0 ); 2862 if( (pIdx = pLevel->plan.u.pIdx)!=0 ){ 2863 if( iCol<0 ){ 2864 sortOrder = 0; 2865 testcase( (pLevel->plan.wsFlags & WHERE_REVERSE)!=0 ); 2866 }else{ 2867 int n = pIdx->nColumn; 2868 for(j=0; j<n; j++){ 2869 if( iCol==pIdx->aiColumn[j] ) break; 2870 } 2871 if( j>=n ) return 0; 2872 sortOrder = pIdx->aSortOrder[j]; 2873 testcase( (pLevel->plan.wsFlags & WHERE_REVERSE)!=0 ); 2874 } 2875 }else{ 2876 if( iCol!=(-1) ) return 0; 2877 sortOrder = 0; 2878 testcase( (pLevel->plan.wsFlags & WHERE_REVERSE)!=0 ); 2879 } 2880 if( (pLevel->plan.wsFlags & WHERE_REVERSE)!=0 ){ 2881 assert( sortOrder==0 || sortOrder==1 ); 2882 testcase( sortOrder==1 ); 2883 sortOrder = 1 - sortOrder; 2884 } 2885 return sortOrder+2; 2886 } 2887 return 0; 2888 } 2889 2890 /* 2891 ** This routine decides if pIdx can be used to satisfy the ORDER BY 2892 ** clause, either in whole or in part. The return value is the 2893 ** cumulative number of terms in the ORDER BY clause that are satisfied 2894 ** by the index pIdx and other indices in outer loops. 2895 ** 2896 ** The table being queried has a cursor number of "base". pIdx is the 2897 ** index that is postulated for use to access the table. 2898 ** 2899 ** The *pbRev value is set to 0 order 1 depending on whether or not 2900 ** pIdx should be run in the forward order or in reverse order. 2901 */ 2902 static int isSortingIndex( 2903 WhereBestIdx *p, /* Best index search context */ 2904 Index *pIdx, /* The index we are testing */ 2905 int base, /* Cursor number for the table to be sorted */ 2906 int *pbRev /* Set to 1 for reverse-order scan of pIdx */ 2907 ){ 2908 int i; /* Number of pIdx terms used */ 2909 int j; /* Number of ORDER BY terms satisfied */ 2910 int sortOrder = 2; /* 0: forward. 1: backward. 2: unknown */ 2911 int nTerm; /* Number of ORDER BY terms */ 2912 struct ExprList_item *pOBItem;/* A term of the ORDER BY clause */ 2913 Table *pTab = pIdx->pTable; /* Table that owns index pIdx */ 2914 ExprList *pOrderBy; /* The ORDER BY clause */ 2915 Parse *pParse = p->pParse; /* Parser context */ 2916 sqlite3 *db = pParse->db; /* Database connection */ 2917 int nPriorSat; /* ORDER BY terms satisfied by outer loops */ 2918 int seenRowid = 0; /* True if an ORDER BY rowid term is seen */ 2919 int uniqueNotNull; /* pIdx is UNIQUE with all terms are NOT NULL */ 2920 2921 if( p->i==0 ){ 2922 nPriorSat = 0; 2923 }else{ 2924 nPriorSat = p->aLevel[p->i-1].plan.nOBSat; 2925 if( (p->aLevel[p->i-1].plan.wsFlags & WHERE_ORDERED)==0 ){ 2926 /* This loop cannot be ordered unless the next outer loop is 2927 ** also ordered */ 2928 return nPriorSat; 2929 } 2930 if( OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ){ 2931 /* Only look at the outer-most loop if the OrderByIdxJoin 2932 ** optimization is disabled */ 2933 return nPriorSat; 2934 } 2935 } 2936 pOrderBy = p->pOrderBy; 2937 assert( pOrderBy!=0 ); 2938 if( pIdx->bUnordered ){ 2939 /* Hash indices (indicated by the "unordered" tag on sqlite_stat1) cannot 2940 ** be used for sorting */ 2941 return nPriorSat; 2942 } 2943 nTerm = pOrderBy->nExpr; 2944 uniqueNotNull = pIdx->onError!=OE_None; 2945 assert( nTerm>0 ); 2946 2947 /* Argument pIdx must either point to a 'real' named index structure, 2948 ** or an index structure allocated on the stack by bestBtreeIndex() to 2949 ** represent the rowid index that is part of every table. */ 2950 assert( pIdx->zName || (pIdx->nColumn==1 && pIdx->aiColumn[0]==-1) ); 2951 2952 /* Match terms of the ORDER BY clause against columns of 2953 ** the index. 2954 ** 2955 ** Note that indices have pIdx->nColumn regular columns plus 2956 ** one additional column containing the rowid. The rowid column 2957 ** of the index is also allowed to match against the ORDER BY 2958 ** clause. 2959 */ 2960 j = nPriorSat; 2961 for(i=0,pOBItem=&pOrderBy->a[j]; j<nTerm && i<=pIdx->nColumn; i++){ 2962 Expr *pOBExpr; /* The expression of the ORDER BY pOBItem */ 2963 CollSeq *pColl; /* The collating sequence of pOBExpr */ 2964 int termSortOrder; /* Sort order for this term */ 2965 int iColumn; /* The i-th column of the index. -1 for rowid */ 2966 int iSortOrder; /* 1 for DESC, 0 for ASC on the i-th index term */ 2967 int isEq; /* Subject to an == or IS NULL constraint */ 2968 int isMatch; /* ORDER BY term matches the index term */ 2969 const char *zColl; /* Name of collating sequence for i-th index term */ 2970 WhereTerm *pConstraint; /* A constraint in the WHERE clause */ 2971 2972 /* If the next term of the ORDER BY clause refers to anything other than 2973 ** a column in the "base" table, then this index will not be of any 2974 ** further use in handling the ORDER BY. */ 2975 pOBExpr = sqlite3ExprSkipCollate(pOBItem->pExpr); 2976 if( pOBExpr->op!=TK_COLUMN || pOBExpr->iTable!=base ){ 2977 break; 2978 } 2979 2980 /* Find column number and collating sequence for the next entry 2981 ** in the index */ 2982 if( pIdx->zName && i<pIdx->nColumn ){ 2983 iColumn = pIdx->aiColumn[i]; 2984 if( iColumn==pIdx->pTable->iPKey ){ 2985 iColumn = -1; 2986 } 2987 iSortOrder = pIdx->aSortOrder[i]; 2988 zColl = pIdx->azColl[i]; 2989 assert( zColl!=0 ); 2990 }else{ 2991 iColumn = -1; 2992 iSortOrder = 0; 2993 zColl = 0; 2994 } 2995 2996 /* Check to see if the column number and collating sequence of the 2997 ** index match the column number and collating sequence of the ORDER BY 2998 ** clause entry. Set isMatch to 1 if they both match. */ 2999 if( pOBExpr->iColumn==iColumn ){ 3000 if( zColl ){ 3001 pColl = sqlite3ExprCollSeq(pParse, pOBItem->pExpr); 3002 if( !pColl ) pColl = db->pDfltColl; 3003 isMatch = sqlite3StrICmp(pColl->zName, zColl)==0; 3004 }else{ 3005 isMatch = 1; 3006 } 3007 }else{ 3008 isMatch = 0; 3009 } 3010 3011 /* termSortOrder is 0 or 1 for whether or not the access loop should 3012 ** run forward or backwards (respectively) in order to satisfy this 3013 ** term of the ORDER BY clause. */ 3014 assert( pOBItem->sortOrder==0 || pOBItem->sortOrder==1 ); 3015 assert( iSortOrder==0 || iSortOrder==1 ); 3016 termSortOrder = iSortOrder ^ pOBItem->sortOrder; 3017 3018 /* If X is the column in the index and ORDER BY clause, check to see 3019 ** if there are any X= or X IS NULL constraints in the WHERE clause. */ 3020 pConstraint = findTerm(p->pWC, base, iColumn, p->notReady, 3021 WO_EQ|WO_ISNULL|WO_IN, pIdx); 3022 if( pConstraint==0 ){ 3023 isEq = 0; 3024 }else if( (pConstraint->eOperator & WO_IN)!=0 ){ 3025 isEq = 0; 3026 }else if( (pConstraint->eOperator & WO_ISNULL)!=0 ){ 3027 uniqueNotNull = 0; 3028 isEq = 1; /* "X IS NULL" means X has only a single value */ 3029 }else if( pConstraint->prereqRight==0 ){ 3030 isEq = 1; /* Constraint "X=constant" means X has only a single value */ 3031 }else{ 3032 Expr *pRight = pConstraint->pExpr->pRight; 3033 if( pRight->op==TK_COLUMN ){ 3034 WHERETRACE((" .. isOrderedColumn(tab=%d,col=%d)", 3035 pRight->iTable, pRight->iColumn)); 3036 isEq = isOrderedColumn(p, pRight->iTable, pRight->iColumn); 3037 WHERETRACE((" -> isEq=%d\n", isEq)); 3038 3039 /* If the constraint is of the form X=Y where Y is an ordered value 3040 ** in an outer loop, then make sure the sort order of Y matches the 3041 ** sort order required for X. */ 3042 if( isMatch && isEq>=2 && isEq!=pOBItem->sortOrder+2 ){ 3043 testcase( isEq==2 ); 3044 testcase( isEq==3 ); 3045 break; 3046 } 3047 }else{ 3048 isEq = 0; /* "X=expr" places no ordering constraints on X */ 3049 } 3050 } 3051 if( !isMatch ){ 3052 if( isEq==0 ){ 3053 break; 3054 }else{ 3055 continue; 3056 } 3057 }else if( isEq!=1 ){ 3058 if( sortOrder==2 ){ 3059 sortOrder = termSortOrder; 3060 }else if( termSortOrder!=sortOrder ){ 3061 break; 3062 } 3063 } 3064 j++; 3065 pOBItem++; 3066 if( iColumn<0 ){ 3067 seenRowid = 1; 3068 break; 3069 }else if( pTab->aCol[iColumn].notNull==0 && isEq!=1 ){ 3070 testcase( isEq==0 ); 3071 testcase( isEq==2 ); 3072 testcase( isEq==3 ); 3073 uniqueNotNull = 0; 3074 } 3075 } 3076 3077 /* If we have not found at least one ORDER BY term that matches the 3078 ** index, then show no progress. */ 3079 if( pOBItem==&pOrderBy->a[nPriorSat] ) return nPriorSat; 3080 3081 /* Return the necessary scan order back to the caller */ 3082 *pbRev = sortOrder & 1; 3083 3084 /* If there was an "ORDER BY rowid" term that matched, or it is only 3085 ** possible for a single row from this table to match, then skip over 3086 ** any additional ORDER BY terms dealing with this table. 3087 */ 3088 if( seenRowid || (uniqueNotNull && i>=pIdx->nColumn) ){ 3089 /* Advance j over additional ORDER BY terms associated with base */ 3090 WhereMaskSet *pMS = p->pWC->pMaskSet; 3091 Bitmask m = ~getMask(pMS, base); 3092 while( j<nTerm && (exprTableUsage(pMS, pOrderBy->a[j].pExpr)&m)==0 ){ 3093 j++; 3094 } 3095 } 3096 return j; 3097 } 3098 3099 /* 3100 ** Find the best query plan for accessing a particular table. Write the 3101 ** best query plan and its cost into the p->cost. 3102 ** 3103 ** The lowest cost plan wins. The cost is an estimate of the amount of 3104 ** CPU and disk I/O needed to process the requested result. 3105 ** Factors that influence cost include: 3106 ** 3107 ** * The estimated number of rows that will be retrieved. (The 3108 ** fewer the better.) 3109 ** 3110 ** * Whether or not sorting must occur. 3111 ** 3112 ** * Whether or not there must be separate lookups in the 3113 ** index and in the main table. 3114 ** 3115 ** If there was an INDEXED BY clause (pSrc->pIndex) attached to the table in 3116 ** the SQL statement, then this function only considers plans using the 3117 ** named index. If no such plan is found, then the returned cost is 3118 ** SQLITE_BIG_DBL. If a plan is found that uses the named index, 3119 ** then the cost is calculated in the usual way. 3120 ** 3121 ** If a NOT INDEXED clause was attached to the table 3122 ** in the SELECT statement, then no indexes are considered. However, the 3123 ** selected plan may still take advantage of the built-in rowid primary key 3124 ** index. 3125 */ 3126 static void bestBtreeIndex(WhereBestIdx *p){ 3127 Parse *pParse = p->pParse; /* The parsing context */ 3128 WhereClause *pWC = p->pWC; /* The WHERE clause */ 3129 struct SrcList_item *pSrc = p->pSrc; /* The FROM clause term to search */ 3130 int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */ 3131 Index *pProbe; /* An index we are evaluating */ 3132 Index *pIdx; /* Copy of pProbe, or zero for IPK index */ 3133 int eqTermMask; /* Current mask of valid equality operators */ 3134 int idxEqTermMask; /* Index mask of valid equality operators */ 3135 Index sPk; /* A fake index object for the primary key */ 3136 tRowcnt aiRowEstPk[2]; /* The aiRowEst[] value for the sPk index */ 3137 int aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 3138 int wsFlagMask; /* Allowed flags in p->cost.plan.wsFlag */ 3139 int nPriorSat; /* ORDER BY terms satisfied by outer loops */ 3140 int nOrderBy; /* Number of ORDER BY terms */ 3141 char bSortInit; /* Initializer for bSort in inner loop */ 3142 char bDistInit; /* Initializer for bDist in inner loop */ 3143 3144 3145 /* Initialize the cost to a worst-case value */ 3146 memset(&p->cost, 0, sizeof(p->cost)); 3147 p->cost.rCost = SQLITE_BIG_DBL; 3148 3149 /* If the pSrc table is the right table of a LEFT JOIN then we may not 3150 ** use an index to satisfy IS NULL constraints on that table. This is 3151 ** because columns might end up being NULL if the table does not match - 3152 ** a circumstance which the index cannot help us discover. Ticket #2177. 3153 */ 3154 if( pSrc->jointype & JT_LEFT ){ 3155 idxEqTermMask = WO_EQ|WO_IN; 3156 }else{ 3157 idxEqTermMask = WO_EQ|WO_IN|WO_ISNULL; 3158 } 3159 3160 if( pSrc->pIndex ){ 3161 /* An INDEXED BY clause specifies a particular index to use */ 3162 pIdx = pProbe = pSrc->pIndex; 3163 wsFlagMask = ~(WHERE_ROWID_EQ|WHERE_ROWID_RANGE); 3164 eqTermMask = idxEqTermMask; 3165 }else{ 3166 /* There is no INDEXED BY clause. Create a fake Index object in local 3167 ** variable sPk to represent the rowid primary key index. Make this 3168 ** fake index the first in a chain of Index objects with all of the real 3169 ** indices to follow */ 3170 Index *pFirst; /* First of real indices on the table */ 3171 memset(&sPk, 0, sizeof(Index)); 3172 sPk.nColumn = 1; 3173 sPk.aiColumn = &aiColumnPk; 3174 sPk.aiRowEst = aiRowEstPk; 3175 sPk.onError = OE_Replace; 3176 sPk.pTable = pSrc->pTab; 3177 aiRowEstPk[0] = pSrc->pTab->nRowEst; 3178 aiRowEstPk[1] = 1; 3179 pFirst = pSrc->pTab->pIndex; 3180 if( pSrc->notIndexed==0 ){ 3181 /* The real indices of the table are only considered if the 3182 ** NOT INDEXED qualifier is omitted from the FROM clause */ 3183 sPk.pNext = pFirst; 3184 } 3185 pProbe = &sPk; 3186 wsFlagMask = ~( 3187 WHERE_COLUMN_IN|WHERE_COLUMN_EQ|WHERE_COLUMN_NULL|WHERE_COLUMN_RANGE 3188 ); 3189 eqTermMask = WO_EQ|WO_IN; 3190 pIdx = 0; 3191 } 3192 3193 nOrderBy = p->pOrderBy ? p->pOrderBy->nExpr : 0; 3194 if( p->i ){ 3195 nPriorSat = p->aLevel[p->i-1].plan.nOBSat; 3196 bSortInit = nPriorSat<nOrderBy; 3197 bDistInit = 0; 3198 }else{ 3199 nPriorSat = 0; 3200 bSortInit = nOrderBy>0; 3201 bDistInit = p->pDistinct!=0; 3202 } 3203 3204 /* Loop over all indices looking for the best one to use 3205 */ 3206 for(; pProbe; pIdx=pProbe=pProbe->pNext){ 3207 const tRowcnt * const aiRowEst = pProbe->aiRowEst; 3208 WhereCost pc; /* Cost of using pProbe */ 3209 double log10N = (double)1; /* base-10 logarithm of nRow (inexact) */ 3210 3211 /* The following variables are populated based on the properties of 3212 ** index being evaluated. They are then used to determine the expected 3213 ** cost and number of rows returned. 3214 ** 3215 ** pc.plan.nEq: 3216 ** Number of equality terms that can be implemented using the index. 3217 ** In other words, the number of initial fields in the index that 3218 ** are used in == or IN or NOT NULL constraints of the WHERE clause. 3219 ** 3220 ** nInMul: 3221 ** The "in-multiplier". This is an estimate of how many seek operations 3222 ** SQLite must perform on the index in question. For example, if the 3223 ** WHERE clause is: 3224 ** 3225 ** WHERE a IN (1, 2, 3) AND b IN (4, 5, 6) 3226 ** 3227 ** SQLite must perform 9 lookups on an index on (a, b), so nInMul is 3228 ** set to 9. Given the same schema and either of the following WHERE 3229 ** clauses: 3230 ** 3231 ** WHERE a = 1 3232 ** WHERE a >= 2 3233 ** 3234 ** nInMul is set to 1. 3235 ** 3236 ** If there exists a WHERE term of the form "x IN (SELECT ...)", then 3237 ** the sub-select is assumed to return 25 rows for the purposes of 3238 ** determining nInMul. 3239 ** 3240 ** bInEst: 3241 ** Set to true if there was at least one "x IN (SELECT ...)" term used 3242 ** in determining the value of nInMul. Note that the RHS of the 3243 ** IN operator must be a SELECT, not a value list, for this variable 3244 ** to be true. 3245 ** 3246 ** rangeDiv: 3247 ** An estimate of a divisor by which to reduce the search space due 3248 ** to inequality constraints. In the absence of sqlite_stat3 ANALYZE 3249 ** data, a single inequality reduces the search space to 1/4rd its 3250 ** original size (rangeDiv==4). Two inequalities reduce the search 3251 ** space to 1/16th of its original size (rangeDiv==16). 3252 ** 3253 ** bSort: 3254 ** Boolean. True if there is an ORDER BY clause that will require an 3255 ** external sort (i.e. scanning the index being evaluated will not 3256 ** correctly order records). 3257 ** 3258 ** bDist: 3259 ** Boolean. True if there is a DISTINCT clause that will require an 3260 ** external btree. 3261 ** 3262 ** bLookup: 3263 ** Boolean. True if a table lookup is required for each index entry 3264 ** visited. In other words, true if this is not a covering index. 3265 ** This is always false for the rowid primary key index of a table. 3266 ** For other indexes, it is true unless all the columns of the table 3267 ** used by the SELECT statement are present in the index (such an 3268 ** index is sometimes described as a covering index). 3269 ** For example, given the index on (a, b), the second of the following 3270 ** two queries requires table b-tree lookups in order to find the value 3271 ** of column c, but the first does not because columns a and b are 3272 ** both available in the index. 3273 ** 3274 ** SELECT a, b FROM tbl WHERE a = 1; 3275 ** SELECT a, b, c FROM tbl WHERE a = 1; 3276 */ 3277 int bInEst = 0; /* True if "x IN (SELECT...)" seen */ 3278 int nInMul = 1; /* Number of distinct equalities to lookup */ 3279 double rangeDiv = (double)1; /* Estimated reduction in search space */ 3280 int nBound = 0; /* Number of range constraints seen */ 3281 char bSort = bSortInit; /* True if external sort required */ 3282 char bDist = bDistInit; /* True if index cannot help with DISTINCT */ 3283 char bLookup = 0; /* True if not a covering index */ 3284 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3285 #ifdef SQLITE_ENABLE_STAT3 3286 WhereTerm *pFirstTerm = 0; /* First term matching the index */ 3287 #endif 3288 3289 WHERETRACE(( 3290 " %s(%s):\n", 3291 pSrc->pTab->zName, (pIdx ? pIdx->zName : "ipk") 3292 )); 3293 memset(&pc, 0, sizeof(pc)); 3294 pc.plan.nOBSat = nPriorSat; 3295 3296 /* Determine the values of pc.plan.nEq and nInMul */ 3297 for(pc.plan.nEq=0; pc.plan.nEq<pProbe->nColumn; pc.plan.nEq++){ 3298 int j = pProbe->aiColumn[pc.plan.nEq]; 3299 pTerm = findTerm(pWC, iCur, j, p->notReady, eqTermMask, pIdx); 3300 if( pTerm==0 ) break; 3301 pc.plan.wsFlags |= (WHERE_COLUMN_EQ|WHERE_ROWID_EQ); 3302 testcase( pTerm->pWC!=pWC ); 3303 if( pTerm->eOperator & WO_IN ){ 3304 Expr *pExpr = pTerm->pExpr; 3305 pc.plan.wsFlags |= WHERE_COLUMN_IN; 3306 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 3307 /* "x IN (SELECT ...)": Assume the SELECT returns 25 rows */ 3308 nInMul *= 25; 3309 bInEst = 1; 3310 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 3311 /* "x IN (value, value, ...)" */ 3312 nInMul *= pExpr->x.pList->nExpr; 3313 } 3314 }else if( pTerm->eOperator & WO_ISNULL ){ 3315 pc.plan.wsFlags |= WHERE_COLUMN_NULL; 3316 } 3317 #ifdef SQLITE_ENABLE_STAT3 3318 if( pc.plan.nEq==0 && pProbe->aSample ) pFirstTerm = pTerm; 3319 #endif 3320 pc.used |= pTerm->prereqRight; 3321 } 3322 3323 /* If the index being considered is UNIQUE, and there is an equality 3324 ** constraint for all columns in the index, then this search will find 3325 ** at most a single row. In this case set the WHERE_UNIQUE flag to 3326 ** indicate this to the caller. 3327 ** 3328 ** Otherwise, if the search may find more than one row, test to see if 3329 ** there is a range constraint on indexed column (pc.plan.nEq+1) that 3330 ** can be optimized using the index. 3331 */ 3332 if( pc.plan.nEq==pProbe->nColumn && pProbe->onError!=OE_None ){ 3333 testcase( pc.plan.wsFlags & WHERE_COLUMN_IN ); 3334 testcase( pc.plan.wsFlags & WHERE_COLUMN_NULL ); 3335 if( (pc.plan.wsFlags & (WHERE_COLUMN_IN|WHERE_COLUMN_NULL))==0 ){ 3336 pc.plan.wsFlags |= WHERE_UNIQUE; 3337 if( p->i==0 || (p->aLevel[p->i-1].plan.wsFlags & WHERE_ALL_UNIQUE)!=0 ){ 3338 pc.plan.wsFlags |= WHERE_ALL_UNIQUE; 3339 } 3340 } 3341 }else if( pProbe->bUnordered==0 ){ 3342 int j; 3343 j = (pc.plan.nEq==pProbe->nColumn ? -1 : pProbe->aiColumn[pc.plan.nEq]); 3344 if( findTerm(pWC, iCur, j, p->notReady, WO_LT|WO_LE|WO_GT|WO_GE, pIdx) ){ 3345 WhereTerm *pTop, *pBtm; 3346 pTop = findTerm(pWC, iCur, j, p->notReady, WO_LT|WO_LE, pIdx); 3347 pBtm = findTerm(pWC, iCur, j, p->notReady, WO_GT|WO_GE, pIdx); 3348 whereRangeScanEst(pParse, pProbe, pc.plan.nEq, pBtm, pTop, &rangeDiv); 3349 if( pTop ){ 3350 nBound = 1; 3351 pc.plan.wsFlags |= WHERE_TOP_LIMIT; 3352 pc.used |= pTop->prereqRight; 3353 testcase( pTop->pWC!=pWC ); 3354 } 3355 if( pBtm ){ 3356 nBound++; 3357 pc.plan.wsFlags |= WHERE_BTM_LIMIT; 3358 pc.used |= pBtm->prereqRight; 3359 testcase( pBtm->pWC!=pWC ); 3360 } 3361 pc.plan.wsFlags |= (WHERE_COLUMN_RANGE|WHERE_ROWID_RANGE); 3362 } 3363 } 3364 3365 /* If there is an ORDER BY clause and the index being considered will 3366 ** naturally scan rows in the required order, set the appropriate flags 3367 ** in pc.plan.wsFlags. Otherwise, if there is an ORDER BY clause but 3368 ** the index will scan rows in a different order, set the bSort 3369 ** variable. */ 3370 if( bSort && (pSrc->jointype & JT_LEFT)==0 ){ 3371 int bRev = 2; 3372 WHERETRACE((" --> before isSortingIndex: nPriorSat=%d\n",nPriorSat)); 3373 pc.plan.nOBSat = isSortingIndex(p, pProbe, iCur, &bRev); 3374 WHERETRACE((" --> after isSortingIndex: bRev=%d nOBSat=%d\n", 3375 bRev, pc.plan.nOBSat)); 3376 if( nPriorSat<pc.plan.nOBSat || (pc.plan.wsFlags & WHERE_ALL_UNIQUE)!=0 ){ 3377 pc.plan.wsFlags |= WHERE_ORDERED; 3378 } 3379 if( nOrderBy==pc.plan.nOBSat ){ 3380 bSort = 0; 3381 pc.plan.wsFlags |= WHERE_ROWID_RANGE|WHERE_COLUMN_RANGE; 3382 } 3383 if( bRev & 1 ) pc.plan.wsFlags |= WHERE_REVERSE; 3384 } 3385 3386 /* If there is a DISTINCT qualifier and this index will scan rows in 3387 ** order of the DISTINCT expressions, clear bDist and set the appropriate 3388 ** flags in pc.plan.wsFlags. */ 3389 if( bDist 3390 && isDistinctIndex(pParse, pWC, pProbe, iCur, p->pDistinct, pc.plan.nEq) 3391 && (pc.plan.wsFlags & WHERE_COLUMN_IN)==0 3392 ){ 3393 bDist = 0; 3394 pc.plan.wsFlags |= WHERE_ROWID_RANGE|WHERE_COLUMN_RANGE|WHERE_DISTINCT; 3395 } 3396 3397 /* If currently calculating the cost of using an index (not the IPK 3398 ** index), determine if all required column data may be obtained without 3399 ** using the main table (i.e. if the index is a covering 3400 ** index for this query). If it is, set the WHERE_IDX_ONLY flag in 3401 ** pc.plan.wsFlags. Otherwise, set the bLookup variable to true. */ 3402 if( pIdx ){ 3403 Bitmask m = pSrc->colUsed; 3404 int j; 3405 for(j=0; j<pIdx->nColumn; j++){ 3406 int x = pIdx->aiColumn[j]; 3407 if( x<BMS-1 ){ 3408 m &= ~(((Bitmask)1)<<x); 3409 } 3410 } 3411 if( m==0 ){ 3412 pc.plan.wsFlags |= WHERE_IDX_ONLY; 3413 }else{ 3414 bLookup = 1; 3415 } 3416 } 3417 3418 /* 3419 ** Estimate the number of rows of output. For an "x IN (SELECT...)" 3420 ** constraint, do not let the estimate exceed half the rows in the table. 3421 */ 3422 pc.plan.nRow = (double)(aiRowEst[pc.plan.nEq] * nInMul); 3423 if( bInEst && pc.plan.nRow*2>aiRowEst[0] ){ 3424 pc.plan.nRow = aiRowEst[0]/2; 3425 nInMul = (int)(pc.plan.nRow / aiRowEst[pc.plan.nEq]); 3426 } 3427 3428 #ifdef SQLITE_ENABLE_STAT3 3429 /* If the constraint is of the form x=VALUE or x IN (E1,E2,...) 3430 ** and we do not think that values of x are unique and if histogram 3431 ** data is available for column x, then it might be possible 3432 ** to get a better estimate on the number of rows based on 3433 ** VALUE and how common that value is according to the histogram. 3434 */ 3435 if( pc.plan.nRow>(double)1 && pc.plan.nEq==1 3436 && pFirstTerm!=0 && aiRowEst[1]>1 ){ 3437 assert( (pFirstTerm->eOperator & (WO_EQ|WO_ISNULL|WO_IN))!=0 ); 3438 if( pFirstTerm->eOperator & (WO_EQ|WO_ISNULL) ){ 3439 testcase( pFirstTerm->eOperator & WO_EQ ); 3440 testcase( pFirstTerm->eOperator & WO_EQUIV ); 3441 testcase( pFirstTerm->eOperator & WO_ISNULL ); 3442 whereEqualScanEst(pParse, pProbe, pFirstTerm->pExpr->pRight, 3443 &pc.plan.nRow); 3444 }else if( bInEst==0 ){ 3445 assert( pFirstTerm->eOperator & WO_IN ); 3446 whereInScanEst(pParse, pProbe, pFirstTerm->pExpr->x.pList, 3447 &pc.plan.nRow); 3448 } 3449 } 3450 #endif /* SQLITE_ENABLE_STAT3 */ 3451 3452 /* Adjust the number of output rows and downward to reflect rows 3453 ** that are excluded by range constraints. 3454 */ 3455 pc.plan.nRow = pc.plan.nRow/rangeDiv; 3456 if( pc.plan.nRow<1 ) pc.plan.nRow = 1; 3457 3458 /* Experiments run on real SQLite databases show that the time needed 3459 ** to do a binary search to locate a row in a table or index is roughly 3460 ** log10(N) times the time to move from one row to the next row within 3461 ** a table or index. The actual times can vary, with the size of 3462 ** records being an important factor. Both moves and searches are 3463 ** slower with larger records, presumably because fewer records fit 3464 ** on one page and hence more pages have to be fetched. 3465 ** 3466 ** The ANALYZE command and the sqlite_stat1 and sqlite_stat3 tables do 3467 ** not give us data on the relative sizes of table and index records. 3468 ** So this computation assumes table records are about twice as big 3469 ** as index records 3470 */ 3471 if( (pc.plan.wsFlags&~(WHERE_REVERSE|WHERE_ORDERED))==WHERE_IDX_ONLY 3472 && (pWC->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 3473 && sqlite3GlobalConfig.bUseCis 3474 && OptimizationEnabled(pParse->db, SQLITE_CoverIdxScan) 3475 ){ 3476 /* This index is not useful for indexing, but it is a covering index. 3477 ** A full-scan of the index might be a little faster than a full-scan 3478 ** of the table, so give this case a cost slightly less than a table 3479 ** scan. */ 3480 pc.rCost = aiRowEst[0]*3 + pProbe->nColumn; 3481 pc.plan.wsFlags |= WHERE_COVER_SCAN|WHERE_COLUMN_RANGE; 3482 }else if( (pc.plan.wsFlags & WHERE_NOT_FULLSCAN)==0 ){ 3483 /* The cost of a full table scan is a number of move operations equal 3484 ** to the number of rows in the table. 3485 ** 3486 ** We add an additional 4x penalty to full table scans. This causes 3487 ** the cost function to err on the side of choosing an index over 3488 ** choosing a full scan. This 4x full-scan penalty is an arguable 3489 ** decision and one which we expect to revisit in the future. But 3490 ** it seems to be working well enough at the moment. 3491 */ 3492 pc.rCost = aiRowEst[0]*4; 3493 pc.plan.wsFlags &= ~WHERE_IDX_ONLY; 3494 if( pIdx ){ 3495 pc.plan.wsFlags &= ~WHERE_ORDERED; 3496 pc.plan.nOBSat = nPriorSat; 3497 } 3498 }else{ 3499 log10N = estLog(aiRowEst[0]); 3500 pc.rCost = pc.plan.nRow; 3501 if( pIdx ){ 3502 if( bLookup ){ 3503 /* For an index lookup followed by a table lookup: 3504 ** nInMul index searches to find the start of each index range 3505 ** + nRow steps through the index 3506 ** + nRow table searches to lookup the table entry using the rowid 3507 */ 3508 pc.rCost += (nInMul + pc.plan.nRow)*log10N; 3509 }else{ 3510 /* For a covering index: 3511 ** nInMul index searches to find the initial entry 3512 ** + nRow steps through the index 3513 */ 3514 pc.rCost += nInMul*log10N; 3515 } 3516 }else{ 3517 /* For a rowid primary key lookup: 3518 ** nInMult table searches to find the initial entry for each range 3519 ** + nRow steps through the table 3520 */ 3521 pc.rCost += nInMul*log10N; 3522 } 3523 } 3524 3525 /* Add in the estimated cost of sorting the result. Actual experimental 3526 ** measurements of sorting performance in SQLite show that sorting time 3527 ** adds C*N*log10(N) to the cost, where N is the number of rows to be 3528 ** sorted and C is a factor between 1.95 and 4.3. We will split the 3529 ** difference and select C of 3.0. 3530 */ 3531 if( bSort ){ 3532 double m = estLog(pc.plan.nRow*(nOrderBy - pc.plan.nOBSat)/nOrderBy); 3533 m *= (double)(pc.plan.nOBSat ? 2 : 3); 3534 pc.rCost += pc.plan.nRow*m; 3535 } 3536 if( bDist ){ 3537 pc.rCost += pc.plan.nRow*estLog(pc.plan.nRow)*3; 3538 } 3539 3540 /**** Cost of using this index has now been computed ****/ 3541 3542 /* If there are additional constraints on this table that cannot 3543 ** be used with the current index, but which might lower the number 3544 ** of output rows, adjust the nRow value accordingly. This only 3545 ** matters if the current index is the least costly, so do not bother 3546 ** with this step if we already know this index will not be chosen. 3547 ** Also, never reduce the output row count below 2 using this step. 3548 ** 3549 ** It is critical that the notValid mask be used here instead of 3550 ** the notReady mask. When computing an "optimal" index, the notReady 3551 ** mask will only have one bit set - the bit for the current table. 3552 ** The notValid mask, on the other hand, always has all bits set for 3553 ** tables that are not in outer loops. If notReady is used here instead 3554 ** of notValid, then a optimal index that depends on inner joins loops 3555 ** might be selected even when there exists an optimal index that has 3556 ** no such dependency. 3557 */ 3558 if( pc.plan.nRow>2 && pc.rCost<=p->cost.rCost ){ 3559 int k; /* Loop counter */ 3560 int nSkipEq = pc.plan.nEq; /* Number of == constraints to skip */ 3561 int nSkipRange = nBound; /* Number of < constraints to skip */ 3562 Bitmask thisTab; /* Bitmap for pSrc */ 3563 3564 thisTab = getMask(pWC->pMaskSet, iCur); 3565 for(pTerm=pWC->a, k=pWC->nTerm; pc.plan.nRow>2 && k; k--, pTerm++){ 3566 if( pTerm->wtFlags & TERM_VIRTUAL ) continue; 3567 if( (pTerm->prereqAll & p->notValid)!=thisTab ) continue; 3568 if( pTerm->eOperator & (WO_EQ|WO_IN|WO_ISNULL) ){ 3569 if( nSkipEq ){ 3570 /* Ignore the first pc.plan.nEq equality matches since the index 3571 ** has already accounted for these */ 3572 nSkipEq--; 3573 }else{ 3574 /* Assume each additional equality match reduces the result 3575 ** set size by a factor of 10 */ 3576 pc.plan.nRow /= 10; 3577 } 3578 }else if( pTerm->eOperator & (WO_LT|WO_LE|WO_GT|WO_GE) ){ 3579 if( nSkipRange ){ 3580 /* Ignore the first nSkipRange range constraints since the index 3581 ** has already accounted for these */ 3582 nSkipRange--; 3583 }else{ 3584 /* Assume each additional range constraint reduces the result 3585 ** set size by a factor of 3. Indexed range constraints reduce 3586 ** the search space by a larger factor: 4. We make indexed range 3587 ** more selective intentionally because of the subjective 3588 ** observation that indexed range constraints really are more 3589 ** selective in practice, on average. */ 3590 pc.plan.nRow /= 3; 3591 } 3592 }else if( (pTerm->eOperator & WO_NOOP)==0 ){ 3593 /* Any other expression lowers the output row count by half */ 3594 pc.plan.nRow /= 2; 3595 } 3596 } 3597 if( pc.plan.nRow<2 ) pc.plan.nRow = 2; 3598 } 3599 3600 3601 WHERETRACE(( 3602 " nEq=%d nInMul=%d rangeDiv=%d bSort=%d bLookup=%d wsFlags=0x%08x\n" 3603 " notReady=0x%llx log10N=%.1f nRow=%.1f cost=%.1f\n" 3604 " used=0x%llx nOBSat=%d\n", 3605 pc.plan.nEq, nInMul, (int)rangeDiv, bSort, bLookup, pc.plan.wsFlags, 3606 p->notReady, log10N, pc.plan.nRow, pc.rCost, pc.used, 3607 pc.plan.nOBSat 3608 )); 3609 3610 /* If this index is the best we have seen so far, then record this 3611 ** index and its cost in the p->cost structure. 3612 */ 3613 if( (!pIdx || pc.plan.wsFlags) && compareCost(&pc, &p->cost) ){ 3614 p->cost = pc; 3615 p->cost.plan.wsFlags &= wsFlagMask; 3616 p->cost.plan.u.pIdx = pIdx; 3617 } 3618 3619 /* If there was an INDEXED BY clause, then only that one index is 3620 ** considered. */ 3621 if( pSrc->pIndex ) break; 3622 3623 /* Reset masks for the next index in the loop */ 3624 wsFlagMask = ~(WHERE_ROWID_EQ|WHERE_ROWID_RANGE); 3625 eqTermMask = idxEqTermMask; 3626 } 3627 3628 /* If there is no ORDER BY clause and the SQLITE_ReverseOrder flag 3629 ** is set, then reverse the order that the index will be scanned 3630 ** in. This is used for application testing, to help find cases 3631 ** where application behavior depends on the (undefined) order that 3632 ** SQLite outputs rows in in the absence of an ORDER BY clause. */ 3633 if( !p->pOrderBy && pParse->db->flags & SQLITE_ReverseOrder ){ 3634 p->cost.plan.wsFlags |= WHERE_REVERSE; 3635 } 3636 3637 assert( p->pOrderBy || (p->cost.plan.wsFlags&WHERE_ORDERED)==0 ); 3638 assert( p->cost.plan.u.pIdx==0 || (p->cost.plan.wsFlags&WHERE_ROWID_EQ)==0 ); 3639 assert( pSrc->pIndex==0 3640 || p->cost.plan.u.pIdx==0 3641 || p->cost.plan.u.pIdx==pSrc->pIndex 3642 ); 3643 3644 WHERETRACE((" best index is %s cost=%.1f\n", 3645 p->cost.plan.u.pIdx ? p->cost.plan.u.pIdx->zName : "ipk", 3646 p->cost.rCost)); 3647 3648 bestOrClauseIndex(p); 3649 bestAutomaticIndex(p); 3650 p->cost.plan.wsFlags |= eqTermMask; 3651 } 3652 3653 /* 3654 ** Find the query plan for accessing table pSrc->pTab. Write the 3655 ** best query plan and its cost into the WhereCost object supplied 3656 ** as the last parameter. This function may calculate the cost of 3657 ** both real and virtual table scans. 3658 ** 3659 ** This function does not take ORDER BY or DISTINCT into account. Nor 3660 ** does it remember the virtual table query plan. All it does is compute 3661 ** the cost while determining if an OR optimization is applicable. The 3662 ** details will be reconsidered later if the optimization is found to be 3663 ** applicable. 3664 */ 3665 static void bestIndex(WhereBestIdx *p){ 3666 #ifndef SQLITE_OMIT_VIRTUALTABLE 3667 if( IsVirtual(p->pSrc->pTab) ){ 3668 sqlite3_index_info *pIdxInfo = 0; 3669 p->ppIdxInfo = &pIdxInfo; 3670 bestVirtualIndex(p); 3671 assert( pIdxInfo!=0 || p->pParse->db->mallocFailed ); 3672 if( pIdxInfo && pIdxInfo->needToFreeIdxStr ){ 3673 sqlite3_free(pIdxInfo->idxStr); 3674 } 3675 sqlite3DbFree(p->pParse->db, pIdxInfo); 3676 }else 3677 #endif 3678 { 3679 bestBtreeIndex(p); 3680 } 3681 } 3682 3683 /* 3684 ** Disable a term in the WHERE clause. Except, do not disable the term 3685 ** if it controls a LEFT OUTER JOIN and it did not originate in the ON 3686 ** or USING clause of that join. 3687 ** 3688 ** Consider the term t2.z='ok' in the following queries: 3689 ** 3690 ** (1) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x WHERE t2.z='ok' 3691 ** (2) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x AND t2.z='ok' 3692 ** (3) SELECT * FROM t1, t2 WHERE t1.a=t2.x AND t2.z='ok' 3693 ** 3694 ** The t2.z='ok' is disabled in the in (2) because it originates 3695 ** in the ON clause. The term is disabled in (3) because it is not part 3696 ** of a LEFT OUTER JOIN. In (1), the term is not disabled. 3697 ** 3698 ** IMPLEMENTATION-OF: R-24597-58655 No tests are done for terms that are 3699 ** completely satisfied by indices. 3700 ** 3701 ** Disabling a term causes that term to not be tested in the inner loop 3702 ** of the join. Disabling is an optimization. When terms are satisfied 3703 ** by indices, we disable them to prevent redundant tests in the inner 3704 ** loop. We would get the correct results if nothing were ever disabled, 3705 ** but joins might run a little slower. The trick is to disable as much 3706 ** as we can without disabling too much. If we disabled in (1), we'd get 3707 ** the wrong answer. See ticket #813. 3708 */ 3709 static void disableTerm(WhereLevel *pLevel, WhereTerm *pTerm){ 3710 if( pTerm 3711 && (pTerm->wtFlags & TERM_CODED)==0 3712 && (pLevel->iLeftJoin==0 || ExprHasProperty(pTerm->pExpr, EP_FromJoin)) 3713 ){ 3714 pTerm->wtFlags |= TERM_CODED; 3715 if( pTerm->iParent>=0 ){ 3716 WhereTerm *pOther = &pTerm->pWC->a[pTerm->iParent]; 3717 if( (--pOther->nChild)==0 ){ 3718 disableTerm(pLevel, pOther); 3719 } 3720 } 3721 } 3722 } 3723 3724 /* 3725 ** Code an OP_Affinity opcode to apply the column affinity string zAff 3726 ** to the n registers starting at base. 3727 ** 3728 ** As an optimization, SQLITE_AFF_NONE entries (which are no-ops) at the 3729 ** beginning and end of zAff are ignored. If all entries in zAff are 3730 ** SQLITE_AFF_NONE, then no code gets generated. 3731 ** 3732 ** This routine makes its own copy of zAff so that the caller is free 3733 ** to modify zAff after this routine returns. 3734 */ 3735 static void codeApplyAffinity(Parse *pParse, int base, int n, char *zAff){ 3736 Vdbe *v = pParse->pVdbe; 3737 if( zAff==0 ){ 3738 assert( pParse->db->mallocFailed ); 3739 return; 3740 } 3741 assert( v!=0 ); 3742 3743 /* Adjust base and n to skip over SQLITE_AFF_NONE entries at the beginning 3744 ** and end of the affinity string. 3745 */ 3746 while( n>0 && zAff[0]==SQLITE_AFF_NONE ){ 3747 n--; 3748 base++; 3749 zAff++; 3750 } 3751 while( n>1 && zAff[n-1]==SQLITE_AFF_NONE ){ 3752 n--; 3753 } 3754 3755 /* Code the OP_Affinity opcode if there is anything left to do. */ 3756 if( n>0 ){ 3757 sqlite3VdbeAddOp2(v, OP_Affinity, base, n); 3758 sqlite3VdbeChangeP4(v, -1, zAff, n); 3759 sqlite3ExprCacheAffinityChange(pParse, base, n); 3760 } 3761 } 3762 3763 3764 /* 3765 ** Generate code for a single equality term of the WHERE clause. An equality 3766 ** term can be either X=expr or X IN (...). pTerm is the term to be 3767 ** coded. 3768 ** 3769 ** The current value for the constraint is left in register iReg. 3770 ** 3771 ** For a constraint of the form X=expr, the expression is evaluated and its 3772 ** result is left on the stack. For constraints of the form X IN (...) 3773 ** this routine sets up a loop that will iterate over all values of X. 3774 */ 3775 static int codeEqualityTerm( 3776 Parse *pParse, /* The parsing context */ 3777 WhereTerm *pTerm, /* The term of the WHERE clause to be coded */ 3778 WhereLevel *pLevel, /* The level of the FROM clause we are working on */ 3779 int iEq, /* Index of the equality term within this level */ 3780 int iTarget /* Attempt to leave results in this register */ 3781 ){ 3782 Expr *pX = pTerm->pExpr; 3783 Vdbe *v = pParse->pVdbe; 3784 int iReg; /* Register holding results */ 3785 3786 assert( iTarget>0 ); 3787 if( pX->op==TK_EQ ){ 3788 iReg = sqlite3ExprCodeTarget(pParse, pX->pRight, iTarget); 3789 }else if( pX->op==TK_ISNULL ){ 3790 iReg = iTarget; 3791 sqlite3VdbeAddOp2(v, OP_Null, 0, iReg); 3792 #ifndef SQLITE_OMIT_SUBQUERY 3793 }else{ 3794 int eType; 3795 int iTab; 3796 struct InLoop *pIn; 3797 u8 bRev = (pLevel->plan.wsFlags & WHERE_REVERSE)!=0; 3798 3799 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 3800 && pLevel->plan.u.pIdx->aSortOrder[iEq] 3801 ){ 3802 testcase( iEq==0 ); 3803 testcase( iEq==pLevel->plan.u.pIdx->nColumn-1 ); 3804 testcase( iEq>0 && iEq+1<pLevel->plan.u.pIdx->nColumn ); 3805 testcase( bRev ); 3806 bRev = !bRev; 3807 } 3808 assert( pX->op==TK_IN ); 3809 iReg = iTarget; 3810 eType = sqlite3FindInIndex(pParse, pX, 0); 3811 if( eType==IN_INDEX_INDEX_DESC ){ 3812 testcase( bRev ); 3813 bRev = !bRev; 3814 } 3815 iTab = pX->iTable; 3816 sqlite3VdbeAddOp2(v, bRev ? OP_Last : OP_Rewind, iTab, 0); 3817 assert( pLevel->plan.wsFlags & WHERE_IN_ABLE ); 3818 if( pLevel->u.in.nIn==0 ){ 3819 pLevel->addrNxt = sqlite3VdbeMakeLabel(v); 3820 } 3821 pLevel->u.in.nIn++; 3822 pLevel->u.in.aInLoop = 3823 sqlite3DbReallocOrFree(pParse->db, pLevel->u.in.aInLoop, 3824 sizeof(pLevel->u.in.aInLoop[0])*pLevel->u.in.nIn); 3825 pIn = pLevel->u.in.aInLoop; 3826 if( pIn ){ 3827 pIn += pLevel->u.in.nIn - 1; 3828 pIn->iCur = iTab; 3829 if( eType==IN_INDEX_ROWID ){ 3830 pIn->addrInTop = sqlite3VdbeAddOp2(v, OP_Rowid, iTab, iReg); 3831 }else{ 3832 pIn->addrInTop = sqlite3VdbeAddOp3(v, OP_Column, iTab, 0, iReg); 3833 } 3834 pIn->eEndLoopOp = bRev ? OP_Prev : OP_Next; 3835 sqlite3VdbeAddOp1(v, OP_IsNull, iReg); 3836 }else{ 3837 pLevel->u.in.nIn = 0; 3838 } 3839 #endif 3840 } 3841 disableTerm(pLevel, pTerm); 3842 return iReg; 3843 } 3844 3845 /* 3846 ** Generate code that will evaluate all == and IN constraints for an 3847 ** index. 3848 ** 3849 ** For example, consider table t1(a,b,c,d,e,f) with index i1(a,b,c). 3850 ** Suppose the WHERE clause is this: a==5 AND b IN (1,2,3) AND c>5 AND c<10 3851 ** The index has as many as three equality constraints, but in this 3852 ** example, the third "c" value is an inequality. So only two 3853 ** constraints are coded. This routine will generate code to evaluate 3854 ** a==5 and b IN (1,2,3). The current values for a and b will be stored 3855 ** in consecutive registers and the index of the first register is returned. 3856 ** 3857 ** In the example above nEq==2. But this subroutine works for any value 3858 ** of nEq including 0. If nEq==0, this routine is nearly a no-op. 3859 ** The only thing it does is allocate the pLevel->iMem memory cell and 3860 ** compute the affinity string. 3861 ** 3862 ** This routine always allocates at least one memory cell and returns 3863 ** the index of that memory cell. The code that 3864 ** calls this routine will use that memory cell to store the termination 3865 ** key value of the loop. If one or more IN operators appear, then 3866 ** this routine allocates an additional nEq memory cells for internal 3867 ** use. 3868 ** 3869 ** Before returning, *pzAff is set to point to a buffer containing a 3870 ** copy of the column affinity string of the index allocated using 3871 ** sqlite3DbMalloc(). Except, entries in the copy of the string associated 3872 ** with equality constraints that use NONE affinity are set to 3873 ** SQLITE_AFF_NONE. This is to deal with SQL such as the following: 3874 ** 3875 ** CREATE TABLE t1(a TEXT PRIMARY KEY, b); 3876 ** SELECT ... FROM t1 AS t2, t1 WHERE t1.a = t2.b; 3877 ** 3878 ** In the example above, the index on t1(a) has TEXT affinity. But since 3879 ** the right hand side of the equality constraint (t2.b) has NONE affinity, 3880 ** no conversion should be attempted before using a t2.b value as part of 3881 ** a key to search the index. Hence the first byte in the returned affinity 3882 ** string in this example would be set to SQLITE_AFF_NONE. 3883 */ 3884 static int codeAllEqualityTerms( 3885 Parse *pParse, /* Parsing context */ 3886 WhereLevel *pLevel, /* Which nested loop of the FROM we are coding */ 3887 WhereClause *pWC, /* The WHERE clause */ 3888 Bitmask notReady, /* Which parts of FROM have not yet been coded */ 3889 int nExtraReg, /* Number of extra registers to allocate */ 3890 char **pzAff /* OUT: Set to point to affinity string */ 3891 ){ 3892 int nEq = pLevel->plan.nEq; /* The number of == or IN constraints to code */ 3893 Vdbe *v = pParse->pVdbe; /* The vm under construction */ 3894 Index *pIdx; /* The index being used for this loop */ 3895 int iCur = pLevel->iTabCur; /* The cursor of the table */ 3896 WhereTerm *pTerm; /* A single constraint term */ 3897 int j; /* Loop counter */ 3898 int regBase; /* Base register */ 3899 int nReg; /* Number of registers to allocate */ 3900 char *zAff; /* Affinity string to return */ 3901 3902 /* This module is only called on query plans that use an index. */ 3903 assert( pLevel->plan.wsFlags & WHERE_INDEXED ); 3904 pIdx = pLevel->plan.u.pIdx; 3905 3906 /* Figure out how many memory cells we will need then allocate them. 3907 */ 3908 regBase = pParse->nMem + 1; 3909 nReg = pLevel->plan.nEq + nExtraReg; 3910 pParse->nMem += nReg; 3911 3912 zAff = sqlite3DbStrDup(pParse->db, sqlite3IndexAffinityStr(v, pIdx)); 3913 if( !zAff ){ 3914 pParse->db->mallocFailed = 1; 3915 } 3916 3917 /* Evaluate the equality constraints 3918 */ 3919 assert( pIdx->nColumn>=nEq ); 3920 for(j=0; j<nEq; j++){ 3921 int r1; 3922 int k = pIdx->aiColumn[j]; 3923 pTerm = findTerm(pWC, iCur, k, notReady, pLevel->plan.wsFlags, pIdx); 3924 if( pTerm==0 ) break; 3925 /* The following true for indices with redundant columns. 3926 ** Ex: CREATE INDEX i1 ON t1(a,b,a); SELECT * FROM t1 WHERE a=0 AND b=0; */ 3927 testcase( (pTerm->wtFlags & TERM_CODED)!=0 ); 3928 testcase( pTerm->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */ 3929 r1 = codeEqualityTerm(pParse, pTerm, pLevel, j, regBase+j); 3930 if( r1!=regBase+j ){ 3931 if( nReg==1 ){ 3932 sqlite3ReleaseTempReg(pParse, regBase); 3933 regBase = r1; 3934 }else{ 3935 sqlite3VdbeAddOp2(v, OP_SCopy, r1, regBase+j); 3936 } 3937 } 3938 testcase( pTerm->eOperator & WO_ISNULL ); 3939 testcase( pTerm->eOperator & WO_IN ); 3940 if( (pTerm->eOperator & (WO_ISNULL|WO_IN))==0 ){ 3941 Expr *pRight = pTerm->pExpr->pRight; 3942 sqlite3ExprCodeIsNullJump(v, pRight, regBase+j, pLevel->addrBrk); 3943 if( zAff ){ 3944 if( sqlite3CompareAffinity(pRight, zAff[j])==SQLITE_AFF_NONE ){ 3945 zAff[j] = SQLITE_AFF_NONE; 3946 } 3947 if( sqlite3ExprNeedsNoAffinityChange(pRight, zAff[j]) ){ 3948 zAff[j] = SQLITE_AFF_NONE; 3949 } 3950 } 3951 } 3952 } 3953 *pzAff = zAff; 3954 return regBase; 3955 } 3956 3957 #ifndef SQLITE_OMIT_EXPLAIN 3958 /* 3959 ** This routine is a helper for explainIndexRange() below 3960 ** 3961 ** pStr holds the text of an expression that we are building up one term 3962 ** at a time. This routine adds a new term to the end of the expression. 3963 ** Terms are separated by AND so add the "AND" text for second and subsequent 3964 ** terms only. 3965 */ 3966 static void explainAppendTerm( 3967 StrAccum *pStr, /* The text expression being built */ 3968 int iTerm, /* Index of this term. First is zero */ 3969 const char *zColumn, /* Name of the column */ 3970 const char *zOp /* Name of the operator */ 3971 ){ 3972 if( iTerm ) sqlite3StrAccumAppend(pStr, " AND ", 5); 3973 sqlite3StrAccumAppend(pStr, zColumn, -1); 3974 sqlite3StrAccumAppend(pStr, zOp, 1); 3975 sqlite3StrAccumAppend(pStr, "?", 1); 3976 } 3977 3978 /* 3979 ** Argument pLevel describes a strategy for scanning table pTab. This 3980 ** function returns a pointer to a string buffer containing a description 3981 ** of the subset of table rows scanned by the strategy in the form of an 3982 ** SQL expression. Or, if all rows are scanned, NULL is returned. 3983 ** 3984 ** For example, if the query: 3985 ** 3986 ** SELECT * FROM t1 WHERE a=1 AND b>2; 3987 ** 3988 ** is run and there is an index on (a, b), then this function returns a 3989 ** string similar to: 3990 ** 3991 ** "a=? AND b>?" 3992 ** 3993 ** The returned pointer points to memory obtained from sqlite3DbMalloc(). 3994 ** It is the responsibility of the caller to free the buffer when it is 3995 ** no longer required. 3996 */ 3997 static char *explainIndexRange(sqlite3 *db, WhereLevel *pLevel, Table *pTab){ 3998 WherePlan *pPlan = &pLevel->plan; 3999 Index *pIndex = pPlan->u.pIdx; 4000 int nEq = pPlan->nEq; 4001 int i, j; 4002 Column *aCol = pTab->aCol; 4003 int *aiColumn = pIndex->aiColumn; 4004 StrAccum txt; 4005 4006 if( nEq==0 && (pPlan->wsFlags & (WHERE_BTM_LIMIT|WHERE_TOP_LIMIT))==0 ){ 4007 return 0; 4008 } 4009 sqlite3StrAccumInit(&txt, 0, 0, SQLITE_MAX_LENGTH); 4010 txt.db = db; 4011 sqlite3StrAccumAppend(&txt, " (", 2); 4012 for(i=0; i<nEq; i++){ 4013 explainAppendTerm(&txt, i, aCol[aiColumn[i]].zName, "="); 4014 } 4015 4016 j = i; 4017 if( pPlan->wsFlags&WHERE_BTM_LIMIT ){ 4018 char *z = (j==pIndex->nColumn ) ? "rowid" : aCol[aiColumn[j]].zName; 4019 explainAppendTerm(&txt, i++, z, ">"); 4020 } 4021 if( pPlan->wsFlags&WHERE_TOP_LIMIT ){ 4022 char *z = (j==pIndex->nColumn ) ? "rowid" : aCol[aiColumn[j]].zName; 4023 explainAppendTerm(&txt, i, z, "<"); 4024 } 4025 sqlite3StrAccumAppend(&txt, ")", 1); 4026 return sqlite3StrAccumFinish(&txt); 4027 } 4028 4029 /* 4030 ** This function is a no-op unless currently processing an EXPLAIN QUERY PLAN 4031 ** command. If the query being compiled is an EXPLAIN QUERY PLAN, a single 4032 ** record is added to the output to describe the table scan strategy in 4033 ** pLevel. 4034 */ 4035 static void explainOneScan( 4036 Parse *pParse, /* Parse context */ 4037 SrcList *pTabList, /* Table list this loop refers to */ 4038 WhereLevel *pLevel, /* Scan to write OP_Explain opcode for */ 4039 int iLevel, /* Value for "level" column of output */ 4040 int iFrom, /* Value for "from" column of output */ 4041 u16 wctrlFlags /* Flags passed to sqlite3WhereBegin() */ 4042 ){ 4043 if( pParse->explain==2 ){ 4044 u32 flags = pLevel->plan.wsFlags; 4045 struct SrcList_item *pItem = &pTabList->a[pLevel->iFrom]; 4046 Vdbe *v = pParse->pVdbe; /* VM being constructed */ 4047 sqlite3 *db = pParse->db; /* Database handle */ 4048 char *zMsg; /* Text to add to EQP output */ 4049 sqlite3_int64 nRow; /* Expected number of rows visited by scan */ 4050 int iId = pParse->iSelectId; /* Select id (left-most output column) */ 4051 int isSearch; /* True for a SEARCH. False for SCAN. */ 4052 4053 if( (flags&WHERE_MULTI_OR) || (wctrlFlags&WHERE_ONETABLE_ONLY) ) return; 4054 4055 isSearch = (pLevel->plan.nEq>0) 4056 || (flags&(WHERE_BTM_LIMIT|WHERE_TOP_LIMIT))!=0 4057 || (wctrlFlags&(WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX)); 4058 4059 zMsg = sqlite3MPrintf(db, "%s", isSearch?"SEARCH":"SCAN"); 4060 if( pItem->pSelect ){ 4061 zMsg = sqlite3MAppendf(db, zMsg, "%s SUBQUERY %d", zMsg,pItem->iSelectId); 4062 }else{ 4063 zMsg = sqlite3MAppendf(db, zMsg, "%s TABLE %s", zMsg, pItem->zName); 4064 } 4065 4066 if( pItem->zAlias ){ 4067 zMsg = sqlite3MAppendf(db, zMsg, "%s AS %s", zMsg, pItem->zAlias); 4068 } 4069 if( (flags & WHERE_INDEXED)!=0 ){ 4070 char *zWhere = explainIndexRange(db, pLevel, pItem->pTab); 4071 zMsg = sqlite3MAppendf(db, zMsg, "%s USING %s%sINDEX%s%s%s", zMsg, 4072 ((flags & WHERE_TEMP_INDEX)?"AUTOMATIC ":""), 4073 ((flags & WHERE_IDX_ONLY)?"COVERING ":""), 4074 ((flags & WHERE_TEMP_INDEX)?"":" "), 4075 ((flags & WHERE_TEMP_INDEX)?"": pLevel->plan.u.pIdx->zName), 4076 zWhere 4077 ); 4078 sqlite3DbFree(db, zWhere); 4079 }else if( flags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){ 4080 zMsg = sqlite3MAppendf(db, zMsg, "%s USING INTEGER PRIMARY KEY", zMsg); 4081 4082 if( flags&WHERE_ROWID_EQ ){ 4083 zMsg = sqlite3MAppendf(db, zMsg, "%s (rowid=?)", zMsg); 4084 }else if( (flags&WHERE_BOTH_LIMIT)==WHERE_BOTH_LIMIT ){ 4085 zMsg = sqlite3MAppendf(db, zMsg, "%s (rowid>? AND rowid<?)", zMsg); 4086 }else if( flags&WHERE_BTM_LIMIT ){ 4087 zMsg = sqlite3MAppendf(db, zMsg, "%s (rowid>?)", zMsg); 4088 }else if( flags&WHERE_TOP_LIMIT ){ 4089 zMsg = sqlite3MAppendf(db, zMsg, "%s (rowid<?)", zMsg); 4090 } 4091 } 4092 #ifndef SQLITE_OMIT_VIRTUALTABLE 4093 else if( (flags & WHERE_VIRTUALTABLE)!=0 ){ 4094 sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx; 4095 zMsg = sqlite3MAppendf(db, zMsg, "%s VIRTUAL TABLE INDEX %d:%s", zMsg, 4096 pVtabIdx->idxNum, pVtabIdx->idxStr); 4097 } 4098 #endif 4099 if( wctrlFlags&(WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX) ){ 4100 testcase( wctrlFlags & WHERE_ORDERBY_MIN ); 4101 nRow = 1; 4102 }else{ 4103 nRow = (sqlite3_int64)pLevel->plan.nRow; 4104 } 4105 zMsg = sqlite3MAppendf(db, zMsg, "%s (~%lld rows)", zMsg, nRow); 4106 sqlite3VdbeAddOp4(v, OP_Explain, iId, iLevel, iFrom, zMsg, P4_DYNAMIC); 4107 } 4108 } 4109 #else 4110 # define explainOneScan(u,v,w,x,y,z) 4111 #endif /* SQLITE_OMIT_EXPLAIN */ 4112 4113 4114 /* 4115 ** Generate code for the start of the iLevel-th loop in the WHERE clause 4116 ** implementation described by pWInfo. 4117 */ 4118 static Bitmask codeOneLoopStart( 4119 WhereInfo *pWInfo, /* Complete information about the WHERE clause */ 4120 int iLevel, /* Which level of pWInfo->a[] should be coded */ 4121 u16 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */ 4122 Bitmask notReady /* Which tables are currently available */ 4123 ){ 4124 int j, k; /* Loop counters */ 4125 int iCur; /* The VDBE cursor for the table */ 4126 int addrNxt; /* Where to jump to continue with the next IN case */ 4127 int omitTable; /* True if we use the index only */ 4128 int bRev; /* True if we need to scan in reverse order */ 4129 WhereLevel *pLevel; /* The where level to be coded */ 4130 WhereClause *pWC; /* Decomposition of the entire WHERE clause */ 4131 WhereTerm *pTerm; /* A WHERE clause term */ 4132 Parse *pParse; /* Parsing context */ 4133 Vdbe *v; /* The prepared stmt under constructions */ 4134 struct SrcList_item *pTabItem; /* FROM clause term being coded */ 4135 int addrBrk; /* Jump here to break out of the loop */ 4136 int addrCont; /* Jump here to continue with next cycle */ 4137 int iRowidReg = 0; /* Rowid is stored in this register, if not zero */ 4138 int iReleaseReg = 0; /* Temp register to free before returning */ 4139 4140 pParse = pWInfo->pParse; 4141 v = pParse->pVdbe; 4142 pWC = pWInfo->pWC; 4143 pLevel = &pWInfo->a[iLevel]; 4144 pTabItem = &pWInfo->pTabList->a[pLevel->iFrom]; 4145 iCur = pTabItem->iCursor; 4146 bRev = (pLevel->plan.wsFlags & WHERE_REVERSE)!=0; 4147 omitTable = (pLevel->plan.wsFlags & WHERE_IDX_ONLY)!=0 4148 && (wctrlFlags & WHERE_FORCE_TABLE)==0; 4149 4150 /* Create labels for the "break" and "continue" instructions 4151 ** for the current loop. Jump to addrBrk to break out of a loop. 4152 ** Jump to cont to go immediately to the next iteration of the 4153 ** loop. 4154 ** 4155 ** When there is an IN operator, we also have a "addrNxt" label that 4156 ** means to continue with the next IN value combination. When 4157 ** there are no IN operators in the constraints, the "addrNxt" label 4158 ** is the same as "addrBrk". 4159 */ 4160 addrBrk = pLevel->addrBrk = pLevel->addrNxt = sqlite3VdbeMakeLabel(v); 4161 addrCont = pLevel->addrCont = sqlite3VdbeMakeLabel(v); 4162 4163 /* If this is the right table of a LEFT OUTER JOIN, allocate and 4164 ** initialize a memory cell that records if this table matches any 4165 ** row of the left table of the join. 4166 */ 4167 if( pLevel->iFrom>0 && (pTabItem[0].jointype & JT_LEFT)!=0 ){ 4168 pLevel->iLeftJoin = ++pParse->nMem; 4169 sqlite3VdbeAddOp2(v, OP_Integer, 0, pLevel->iLeftJoin); 4170 VdbeComment((v, "init LEFT JOIN no-match flag")); 4171 } 4172 4173 /* Special case of a FROM clause subquery implemented as a co-routine */ 4174 if( pTabItem->viaCoroutine ){ 4175 int regYield = pTabItem->regReturn; 4176 sqlite3VdbeAddOp2(v, OP_Integer, pTabItem->addrFillSub-1, regYield); 4177 pLevel->p2 = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 4178 VdbeComment((v, "next row of co-routine %s", pTabItem->pTab->zName)); 4179 sqlite3VdbeAddOp2(v, OP_If, regYield+1, addrBrk); 4180 pLevel->op = OP_Goto; 4181 }else 4182 4183 #ifndef SQLITE_OMIT_VIRTUALTABLE 4184 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 4185 /* Case 0: The table is a virtual-table. Use the VFilter and VNext 4186 ** to access the data. 4187 */ 4188 int iReg; /* P3 Value for OP_VFilter */ 4189 int addrNotFound; 4190 sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx; 4191 int nConstraint = pVtabIdx->nConstraint; 4192 struct sqlite3_index_constraint_usage *aUsage = 4193 pVtabIdx->aConstraintUsage; 4194 const struct sqlite3_index_constraint *aConstraint = 4195 pVtabIdx->aConstraint; 4196 4197 sqlite3ExprCachePush(pParse); 4198 iReg = sqlite3GetTempRange(pParse, nConstraint+2); 4199 addrNotFound = pLevel->addrBrk; 4200 for(j=1; j<=nConstraint; j++){ 4201 for(k=0; k<nConstraint; k++){ 4202 if( aUsage[k].argvIndex==j ){ 4203 int iTarget = iReg+j+1; 4204 pTerm = &pWC->a[aConstraint[k].iTermOffset]; 4205 if( pTerm->eOperator & WO_IN ){ 4206 codeEqualityTerm(pParse, pTerm, pLevel, k, iTarget); 4207 addrNotFound = pLevel->addrNxt; 4208 }else{ 4209 sqlite3ExprCode(pParse, pTerm->pExpr->pRight, iTarget); 4210 } 4211 break; 4212 } 4213 } 4214 if( k==nConstraint ) break; 4215 } 4216 sqlite3VdbeAddOp2(v, OP_Integer, pVtabIdx->idxNum, iReg); 4217 sqlite3VdbeAddOp2(v, OP_Integer, j-1, iReg+1); 4218 sqlite3VdbeAddOp4(v, OP_VFilter, iCur, addrNotFound, iReg, pVtabIdx->idxStr, 4219 pVtabIdx->needToFreeIdxStr ? P4_MPRINTF : P4_STATIC); 4220 pVtabIdx->needToFreeIdxStr = 0; 4221 for(j=0; j<nConstraint; j++){ 4222 if( aUsage[j].omit ){ 4223 int iTerm = aConstraint[j].iTermOffset; 4224 disableTerm(pLevel, &pWC->a[iTerm]); 4225 } 4226 } 4227 pLevel->op = OP_VNext; 4228 pLevel->p1 = iCur; 4229 pLevel->p2 = sqlite3VdbeCurrentAddr(v); 4230 sqlite3ReleaseTempRange(pParse, iReg, nConstraint+2); 4231 sqlite3ExprCachePop(pParse, 1); 4232 }else 4233 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 4234 4235 if( pLevel->plan.wsFlags & WHERE_ROWID_EQ ){ 4236 /* Case 1: We can directly reference a single row using an 4237 ** equality comparison against the ROWID field. Or 4238 ** we reference multiple rows using a "rowid IN (...)" 4239 ** construct. 4240 */ 4241 iReleaseReg = sqlite3GetTempReg(pParse); 4242 pTerm = findTerm(pWC, iCur, -1, notReady, WO_EQ|WO_IN, 0); 4243 assert( pTerm!=0 ); 4244 assert( pTerm->pExpr!=0 ); 4245 assert( omitTable==0 ); 4246 testcase( pTerm->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */ 4247 iRowidReg = codeEqualityTerm(pParse, pTerm, pLevel, 0, iReleaseReg); 4248 addrNxt = pLevel->addrNxt; 4249 sqlite3VdbeAddOp2(v, OP_MustBeInt, iRowidReg, addrNxt); 4250 sqlite3VdbeAddOp3(v, OP_NotExists, iCur, addrNxt, iRowidReg); 4251 sqlite3ExprCacheAffinityChange(pParse, iRowidReg, 1); 4252 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); 4253 VdbeComment((v, "pk")); 4254 pLevel->op = OP_Noop; 4255 }else if( pLevel->plan.wsFlags & WHERE_ROWID_RANGE ){ 4256 /* Case 2: We have an inequality comparison against the ROWID field. 4257 */ 4258 int testOp = OP_Noop; 4259 int start; 4260 int memEndValue = 0; 4261 WhereTerm *pStart, *pEnd; 4262 4263 assert( omitTable==0 ); 4264 pStart = findTerm(pWC, iCur, -1, notReady, WO_GT|WO_GE, 0); 4265 pEnd = findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE, 0); 4266 if( bRev ){ 4267 pTerm = pStart; 4268 pStart = pEnd; 4269 pEnd = pTerm; 4270 } 4271 if( pStart ){ 4272 Expr *pX; /* The expression that defines the start bound */ 4273 int r1, rTemp; /* Registers for holding the start boundary */ 4274 4275 /* The following constant maps TK_xx codes into corresponding 4276 ** seek opcodes. It depends on a particular ordering of TK_xx 4277 */ 4278 const u8 aMoveOp[] = { 4279 /* TK_GT */ OP_SeekGt, 4280 /* TK_LE */ OP_SeekLe, 4281 /* TK_LT */ OP_SeekLt, 4282 /* TK_GE */ OP_SeekGe 4283 }; 4284 assert( TK_LE==TK_GT+1 ); /* Make sure the ordering.. */ 4285 assert( TK_LT==TK_GT+2 ); /* ... of the TK_xx values... */ 4286 assert( TK_GE==TK_GT+3 ); /* ... is correcct. */ 4287 4288 testcase( pStart->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */ 4289 pX = pStart->pExpr; 4290 assert( pX!=0 ); 4291 assert( pStart->leftCursor==iCur ); 4292 r1 = sqlite3ExprCodeTemp(pParse, pX->pRight, &rTemp); 4293 sqlite3VdbeAddOp3(v, aMoveOp[pX->op-TK_GT], iCur, addrBrk, r1); 4294 VdbeComment((v, "pk")); 4295 sqlite3ExprCacheAffinityChange(pParse, r1, 1); 4296 sqlite3ReleaseTempReg(pParse, rTemp); 4297 disableTerm(pLevel, pStart); 4298 }else{ 4299 sqlite3VdbeAddOp2(v, bRev ? OP_Last : OP_Rewind, iCur, addrBrk); 4300 } 4301 if( pEnd ){ 4302 Expr *pX; 4303 pX = pEnd->pExpr; 4304 assert( pX!=0 ); 4305 assert( pEnd->leftCursor==iCur ); 4306 testcase( pEnd->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */ 4307 memEndValue = ++pParse->nMem; 4308 sqlite3ExprCode(pParse, pX->pRight, memEndValue); 4309 if( pX->op==TK_LT || pX->op==TK_GT ){ 4310 testOp = bRev ? OP_Le : OP_Ge; 4311 }else{ 4312 testOp = bRev ? OP_Lt : OP_Gt; 4313 } 4314 disableTerm(pLevel, pEnd); 4315 } 4316 start = sqlite3VdbeCurrentAddr(v); 4317 pLevel->op = bRev ? OP_Prev : OP_Next; 4318 pLevel->p1 = iCur; 4319 pLevel->p2 = start; 4320 if( pStart==0 && pEnd==0 ){ 4321 pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP; 4322 }else{ 4323 assert( pLevel->p5==0 ); 4324 } 4325 if( testOp!=OP_Noop ){ 4326 iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse); 4327 sqlite3VdbeAddOp2(v, OP_Rowid, iCur, iRowidReg); 4328 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); 4329 sqlite3VdbeAddOp3(v, testOp, memEndValue, addrBrk, iRowidReg); 4330 sqlite3VdbeChangeP5(v, SQLITE_AFF_NUMERIC | SQLITE_JUMPIFNULL); 4331 } 4332 }else if( pLevel->plan.wsFlags & (WHERE_COLUMN_RANGE|WHERE_COLUMN_EQ) ){ 4333 /* Case 3: A scan using an index. 4334 ** 4335 ** The WHERE clause may contain zero or more equality 4336 ** terms ("==" or "IN" operators) that refer to the N 4337 ** left-most columns of the index. It may also contain 4338 ** inequality constraints (>, <, >= or <=) on the indexed 4339 ** column that immediately follows the N equalities. Only 4340 ** the right-most column can be an inequality - the rest must 4341 ** use the "==" and "IN" operators. For example, if the 4342 ** index is on (x,y,z), then the following clauses are all 4343 ** optimized: 4344 ** 4345 ** x=5 4346 ** x=5 AND y=10 4347 ** x=5 AND y<10 4348 ** x=5 AND y>5 AND y<10 4349 ** x=5 AND y=5 AND z<=10 4350 ** 4351 ** The z<10 term of the following cannot be used, only 4352 ** the x=5 term: 4353 ** 4354 ** x=5 AND z<10 4355 ** 4356 ** N may be zero if there are inequality constraints. 4357 ** If there are no inequality constraints, then N is at 4358 ** least one. 4359 ** 4360 ** This case is also used when there are no WHERE clause 4361 ** constraints but an index is selected anyway, in order 4362 ** to force the output order to conform to an ORDER BY. 4363 */ 4364 static const u8 aStartOp[] = { 4365 0, 4366 0, 4367 OP_Rewind, /* 2: (!start_constraints && startEq && !bRev) */ 4368 OP_Last, /* 3: (!start_constraints && startEq && bRev) */ 4369 OP_SeekGt, /* 4: (start_constraints && !startEq && !bRev) */ 4370 OP_SeekLt, /* 5: (start_constraints && !startEq && bRev) */ 4371 OP_SeekGe, /* 6: (start_constraints && startEq && !bRev) */ 4372 OP_SeekLe /* 7: (start_constraints && startEq && bRev) */ 4373 }; 4374 static const u8 aEndOp[] = { 4375 OP_Noop, /* 0: (!end_constraints) */ 4376 OP_IdxGE, /* 1: (end_constraints && !bRev) */ 4377 OP_IdxLT /* 2: (end_constraints && bRev) */ 4378 }; 4379 int nEq = pLevel->plan.nEq; /* Number of == or IN terms */ 4380 int isMinQuery = 0; /* If this is an optimized SELECT min(x).. */ 4381 int regBase; /* Base register holding constraint values */ 4382 int r1; /* Temp register */ 4383 WhereTerm *pRangeStart = 0; /* Inequality constraint at range start */ 4384 WhereTerm *pRangeEnd = 0; /* Inequality constraint at range end */ 4385 int startEq; /* True if range start uses ==, >= or <= */ 4386 int endEq; /* True if range end uses ==, >= or <= */ 4387 int start_constraints; /* Start of range is constrained */ 4388 int nConstraint; /* Number of constraint terms */ 4389 Index *pIdx; /* The index we will be using */ 4390 int iIdxCur; /* The VDBE cursor for the index */ 4391 int nExtraReg = 0; /* Number of extra registers needed */ 4392 int op; /* Instruction opcode */ 4393 char *zStartAff; /* Affinity for start of range constraint */ 4394 char *zEndAff; /* Affinity for end of range constraint */ 4395 4396 pIdx = pLevel->plan.u.pIdx; 4397 iIdxCur = pLevel->iIdxCur; 4398 k = (nEq==pIdx->nColumn ? -1 : pIdx->aiColumn[nEq]); 4399 4400 /* If this loop satisfies a sort order (pOrderBy) request that 4401 ** was passed to this function to implement a "SELECT min(x) ..." 4402 ** query, then the caller will only allow the loop to run for 4403 ** a single iteration. This means that the first row returned 4404 ** should not have a NULL value stored in 'x'. If column 'x' is 4405 ** the first one after the nEq equality constraints in the index, 4406 ** this requires some special handling. 4407 */ 4408 if( (wctrlFlags&WHERE_ORDERBY_MIN)!=0 4409 && (pLevel->plan.wsFlags&WHERE_ORDERED) 4410 && (pIdx->nColumn>nEq) 4411 ){ 4412 /* assert( pOrderBy->nExpr==1 ); */ 4413 /* assert( pOrderBy->a[0].pExpr->iColumn==pIdx->aiColumn[nEq] ); */ 4414 isMinQuery = 1; 4415 nExtraReg = 1; 4416 } 4417 4418 /* Find any inequality constraint terms for the start and end 4419 ** of the range. 4420 */ 4421 if( pLevel->plan.wsFlags & WHERE_TOP_LIMIT ){ 4422 pRangeEnd = findTerm(pWC, iCur, k, notReady, (WO_LT|WO_LE), pIdx); 4423 nExtraReg = 1; 4424 } 4425 if( pLevel->plan.wsFlags & WHERE_BTM_LIMIT ){ 4426 pRangeStart = findTerm(pWC, iCur, k, notReady, (WO_GT|WO_GE), pIdx); 4427 nExtraReg = 1; 4428 } 4429 4430 /* Generate code to evaluate all constraint terms using == or IN 4431 ** and store the values of those terms in an array of registers 4432 ** starting at regBase. 4433 */ 4434 regBase = codeAllEqualityTerms( 4435 pParse, pLevel, pWC, notReady, nExtraReg, &zStartAff 4436 ); 4437 zEndAff = sqlite3DbStrDup(pParse->db, zStartAff); 4438 addrNxt = pLevel->addrNxt; 4439 4440 /* If we are doing a reverse order scan on an ascending index, or 4441 ** a forward order scan on a descending index, interchange the 4442 ** start and end terms (pRangeStart and pRangeEnd). 4443 */ 4444 if( (nEq<pIdx->nColumn && bRev==(pIdx->aSortOrder[nEq]==SQLITE_SO_ASC)) 4445 || (bRev && pIdx->nColumn==nEq) 4446 ){ 4447 SWAP(WhereTerm *, pRangeEnd, pRangeStart); 4448 } 4449 4450 testcase( pRangeStart && pRangeStart->eOperator & WO_LE ); 4451 testcase( pRangeStart && pRangeStart->eOperator & WO_GE ); 4452 testcase( pRangeEnd && pRangeEnd->eOperator & WO_LE ); 4453 testcase( pRangeEnd && pRangeEnd->eOperator & WO_GE ); 4454 startEq = !pRangeStart || pRangeStart->eOperator & (WO_LE|WO_GE); 4455 endEq = !pRangeEnd || pRangeEnd->eOperator & (WO_LE|WO_GE); 4456 start_constraints = pRangeStart || nEq>0; 4457 4458 /* Seek the index cursor to the start of the range. */ 4459 nConstraint = nEq; 4460 if( pRangeStart ){ 4461 Expr *pRight = pRangeStart->pExpr->pRight; 4462 sqlite3ExprCode(pParse, pRight, regBase+nEq); 4463 if( (pRangeStart->wtFlags & TERM_VNULL)==0 ){ 4464 sqlite3ExprCodeIsNullJump(v, pRight, regBase+nEq, addrNxt); 4465 } 4466 if( zStartAff ){ 4467 if( sqlite3CompareAffinity(pRight, zStartAff[nEq])==SQLITE_AFF_NONE){ 4468 /* Since the comparison is to be performed with no conversions 4469 ** applied to the operands, set the affinity to apply to pRight to 4470 ** SQLITE_AFF_NONE. */ 4471 zStartAff[nEq] = SQLITE_AFF_NONE; 4472 } 4473 if( sqlite3ExprNeedsNoAffinityChange(pRight, zStartAff[nEq]) ){ 4474 zStartAff[nEq] = SQLITE_AFF_NONE; 4475 } 4476 } 4477 nConstraint++; 4478 testcase( pRangeStart->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */ 4479 }else if( isMinQuery ){ 4480 sqlite3VdbeAddOp2(v, OP_Null, 0, regBase+nEq); 4481 nConstraint++; 4482 startEq = 0; 4483 start_constraints = 1; 4484 } 4485 codeApplyAffinity(pParse, regBase, nConstraint, zStartAff); 4486 op = aStartOp[(start_constraints<<2) + (startEq<<1) + bRev]; 4487 assert( op!=0 ); 4488 testcase( op==OP_Rewind ); 4489 testcase( op==OP_Last ); 4490 testcase( op==OP_SeekGt ); 4491 testcase( op==OP_SeekGe ); 4492 testcase( op==OP_SeekLe ); 4493 testcase( op==OP_SeekLt ); 4494 sqlite3VdbeAddOp4Int(v, op, iIdxCur, addrNxt, regBase, nConstraint); 4495 4496 /* Load the value for the inequality constraint at the end of the 4497 ** range (if any). 4498 */ 4499 nConstraint = nEq; 4500 if( pRangeEnd ){ 4501 Expr *pRight = pRangeEnd->pExpr->pRight; 4502 sqlite3ExprCacheRemove(pParse, regBase+nEq, 1); 4503 sqlite3ExprCode(pParse, pRight, regBase+nEq); 4504 if( (pRangeEnd->wtFlags & TERM_VNULL)==0 ){ 4505 sqlite3ExprCodeIsNullJump(v, pRight, regBase+nEq, addrNxt); 4506 } 4507 if( zEndAff ){ 4508 if( sqlite3CompareAffinity(pRight, zEndAff[nEq])==SQLITE_AFF_NONE){ 4509 /* Since the comparison is to be performed with no conversions 4510 ** applied to the operands, set the affinity to apply to pRight to 4511 ** SQLITE_AFF_NONE. */ 4512 zEndAff[nEq] = SQLITE_AFF_NONE; 4513 } 4514 if( sqlite3ExprNeedsNoAffinityChange(pRight, zEndAff[nEq]) ){ 4515 zEndAff[nEq] = SQLITE_AFF_NONE; 4516 } 4517 } 4518 codeApplyAffinity(pParse, regBase, nEq+1, zEndAff); 4519 nConstraint++; 4520 testcase( pRangeEnd->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */ 4521 } 4522 sqlite3DbFree(pParse->db, zStartAff); 4523 sqlite3DbFree(pParse->db, zEndAff); 4524 4525 /* Top of the loop body */ 4526 pLevel->p2 = sqlite3VdbeCurrentAddr(v); 4527 4528 /* Check if the index cursor is past the end of the range. */ 4529 op = aEndOp[(pRangeEnd || nEq) * (1 + bRev)]; 4530 testcase( op==OP_Noop ); 4531 testcase( op==OP_IdxGE ); 4532 testcase( op==OP_IdxLT ); 4533 if( op!=OP_Noop ){ 4534 sqlite3VdbeAddOp4Int(v, op, iIdxCur, addrNxt, regBase, nConstraint); 4535 sqlite3VdbeChangeP5(v, endEq!=bRev ?1:0); 4536 } 4537 4538 /* If there are inequality constraints, check that the value 4539 ** of the table column that the inequality contrains is not NULL. 4540 ** If it is, jump to the next iteration of the loop. 4541 */ 4542 r1 = sqlite3GetTempReg(pParse); 4543 testcase( pLevel->plan.wsFlags & WHERE_BTM_LIMIT ); 4544 testcase( pLevel->plan.wsFlags & WHERE_TOP_LIMIT ); 4545 if( (pLevel->plan.wsFlags & (WHERE_BTM_LIMIT|WHERE_TOP_LIMIT))!=0 ){ 4546 sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, nEq, r1); 4547 sqlite3VdbeAddOp2(v, OP_IsNull, r1, addrCont); 4548 } 4549 sqlite3ReleaseTempReg(pParse, r1); 4550 4551 /* Seek the table cursor, if required */ 4552 disableTerm(pLevel, pRangeStart); 4553 disableTerm(pLevel, pRangeEnd); 4554 if( !omitTable ){ 4555 iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse); 4556 sqlite3VdbeAddOp2(v, OP_IdxRowid, iIdxCur, iRowidReg); 4557 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); 4558 sqlite3VdbeAddOp2(v, OP_Seek, iCur, iRowidReg); /* Deferred seek */ 4559 } 4560 4561 /* Record the instruction used to terminate the loop. Disable 4562 ** WHERE clause terms made redundant by the index range scan. 4563 */ 4564 if( pLevel->plan.wsFlags & WHERE_UNIQUE ){ 4565 pLevel->op = OP_Noop; 4566 }else if( bRev ){ 4567 pLevel->op = OP_Prev; 4568 }else{ 4569 pLevel->op = OP_Next; 4570 } 4571 pLevel->p1 = iIdxCur; 4572 if( pLevel->plan.wsFlags & WHERE_COVER_SCAN ){ 4573 pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP; 4574 }else{ 4575 assert( pLevel->p5==0 ); 4576 } 4577 }else 4578 4579 #ifndef SQLITE_OMIT_OR_OPTIMIZATION 4580 if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){ 4581 /* Case 4: Two or more separately indexed terms connected by OR 4582 ** 4583 ** Example: 4584 ** 4585 ** CREATE TABLE t1(a,b,c,d); 4586 ** CREATE INDEX i1 ON t1(a); 4587 ** CREATE INDEX i2 ON t1(b); 4588 ** CREATE INDEX i3 ON t1(c); 4589 ** 4590 ** SELECT * FROM t1 WHERE a=5 OR b=7 OR (c=11 AND d=13) 4591 ** 4592 ** In the example, there are three indexed terms connected by OR. 4593 ** The top of the loop looks like this: 4594 ** 4595 ** Null 1 # Zero the rowset in reg 1 4596 ** 4597 ** Then, for each indexed term, the following. The arguments to 4598 ** RowSetTest are such that the rowid of the current row is inserted 4599 ** into the RowSet. If it is already present, control skips the 4600 ** Gosub opcode and jumps straight to the code generated by WhereEnd(). 4601 ** 4602 ** sqlite3WhereBegin(<term>) 4603 ** RowSetTest # Insert rowid into rowset 4604 ** Gosub 2 A 4605 ** sqlite3WhereEnd() 4606 ** 4607 ** Following the above, code to terminate the loop. Label A, the target 4608 ** of the Gosub above, jumps to the instruction right after the Goto. 4609 ** 4610 ** Null 1 # Zero the rowset in reg 1 4611 ** Goto B # The loop is finished. 4612 ** 4613 ** A: <loop body> # Return data, whatever. 4614 ** 4615 ** Return 2 # Jump back to the Gosub 4616 ** 4617 ** B: <after the loop> 4618 ** 4619 */ 4620 WhereClause *pOrWc; /* The OR-clause broken out into subterms */ 4621 SrcList *pOrTab; /* Shortened table list or OR-clause generation */ 4622 Index *pCov = 0; /* Potential covering index (or NULL) */ 4623 int iCovCur = pParse->nTab++; /* Cursor used for index scans (if any) */ 4624 4625 int regReturn = ++pParse->nMem; /* Register used with OP_Gosub */ 4626 int regRowset = 0; /* Register for RowSet object */ 4627 int regRowid = 0; /* Register holding rowid */ 4628 int iLoopBody = sqlite3VdbeMakeLabel(v); /* Start of loop body */ 4629 int iRetInit; /* Address of regReturn init */ 4630 int untestedTerms = 0; /* Some terms not completely tested */ 4631 int ii; /* Loop counter */ 4632 Expr *pAndExpr = 0; /* An ".. AND (...)" expression */ 4633 4634 pTerm = pLevel->plan.u.pTerm; 4635 assert( pTerm!=0 ); 4636 assert( pTerm->eOperator & WO_OR ); 4637 assert( (pTerm->wtFlags & TERM_ORINFO)!=0 ); 4638 pOrWc = &pTerm->u.pOrInfo->wc; 4639 pLevel->op = OP_Return; 4640 pLevel->p1 = regReturn; 4641 4642 /* Set up a new SrcList in pOrTab containing the table being scanned 4643 ** by this loop in the a[0] slot and all notReady tables in a[1..] slots. 4644 ** This becomes the SrcList in the recursive call to sqlite3WhereBegin(). 4645 */ 4646 if( pWInfo->nLevel>1 ){ 4647 int nNotReady; /* The number of notReady tables */ 4648 struct SrcList_item *origSrc; /* Original list of tables */ 4649 nNotReady = pWInfo->nLevel - iLevel - 1; 4650 pOrTab = sqlite3StackAllocRaw(pParse->db, 4651 sizeof(*pOrTab)+ nNotReady*sizeof(pOrTab->a[0])); 4652 if( pOrTab==0 ) return notReady; 4653 pOrTab->nAlloc = (i16)(nNotReady + 1); 4654 pOrTab->nSrc = pOrTab->nAlloc; 4655 memcpy(pOrTab->a, pTabItem, sizeof(*pTabItem)); 4656 origSrc = pWInfo->pTabList->a; 4657 for(k=1; k<=nNotReady; k++){ 4658 memcpy(&pOrTab->a[k], &origSrc[pLevel[k].iFrom], sizeof(pOrTab->a[k])); 4659 } 4660 }else{ 4661 pOrTab = pWInfo->pTabList; 4662 } 4663 4664 /* Initialize the rowset register to contain NULL. An SQL NULL is 4665 ** equivalent to an empty rowset. 4666 ** 4667 ** Also initialize regReturn to contain the address of the instruction 4668 ** immediately following the OP_Return at the bottom of the loop. This 4669 ** is required in a few obscure LEFT JOIN cases where control jumps 4670 ** over the top of the loop into the body of it. In this case the 4671 ** correct response for the end-of-loop code (the OP_Return) is to 4672 ** fall through to the next instruction, just as an OP_Next does if 4673 ** called on an uninitialized cursor. 4674 */ 4675 if( (wctrlFlags & WHERE_DUPLICATES_OK)==0 ){ 4676 regRowset = ++pParse->nMem; 4677 regRowid = ++pParse->nMem; 4678 sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset); 4679 } 4680 iRetInit = sqlite3VdbeAddOp2(v, OP_Integer, 0, regReturn); 4681 4682 /* If the original WHERE clause is z of the form: (x1 OR x2 OR ...) AND y 4683 ** Then for every term xN, evaluate as the subexpression: xN AND z 4684 ** That way, terms in y that are factored into the disjunction will 4685 ** be picked up by the recursive calls to sqlite3WhereBegin() below. 4686 ** 4687 ** Actually, each subexpression is converted to "xN AND w" where w is 4688 ** the "interesting" terms of z - terms that did not originate in the 4689 ** ON or USING clause of a LEFT JOIN, and terms that are usable as 4690 ** indices. 4691 */ 4692 if( pWC->nTerm>1 ){ 4693 int iTerm; 4694 for(iTerm=0; iTerm<pWC->nTerm; iTerm++){ 4695 Expr *pExpr = pWC->a[iTerm].pExpr; 4696 if( ExprHasProperty(pExpr, EP_FromJoin) ) continue; 4697 if( pWC->a[iTerm].wtFlags & (TERM_VIRTUAL|TERM_ORINFO) ) continue; 4698 if( (pWC->a[iTerm].eOperator & WO_ALL)==0 ) continue; 4699 pExpr = sqlite3ExprDup(pParse->db, pExpr, 0); 4700 pAndExpr = sqlite3ExprAnd(pParse->db, pAndExpr, pExpr); 4701 } 4702 if( pAndExpr ){ 4703 pAndExpr = sqlite3PExpr(pParse, TK_AND, 0, pAndExpr, 0); 4704 } 4705 } 4706 4707 for(ii=0; ii<pOrWc->nTerm; ii++){ 4708 WhereTerm *pOrTerm = &pOrWc->a[ii]; 4709 if( pOrTerm->leftCursor==iCur || (pOrTerm->eOperator & WO_AND)!=0 ){ 4710 WhereInfo *pSubWInfo; /* Info for single OR-term scan */ 4711 Expr *pOrExpr = pOrTerm->pExpr; 4712 if( pAndExpr ){ 4713 pAndExpr->pLeft = pOrExpr; 4714 pOrExpr = pAndExpr; 4715 } 4716 /* Loop through table entries that match term pOrTerm. */ 4717 pSubWInfo = sqlite3WhereBegin(pParse, pOrTab, pOrExpr, 0, 0, 4718 WHERE_OMIT_OPEN_CLOSE | WHERE_AND_ONLY | 4719 WHERE_FORCE_TABLE | WHERE_ONETABLE_ONLY, iCovCur); 4720 assert( pSubWInfo || pParse->nErr || pParse->db->mallocFailed ); 4721 if( pSubWInfo ){ 4722 WhereLevel *pLvl; 4723 explainOneScan( 4724 pParse, pOrTab, &pSubWInfo->a[0], iLevel, pLevel->iFrom, 0 4725 ); 4726 if( (wctrlFlags & WHERE_DUPLICATES_OK)==0 ){ 4727 int iSet = ((ii==pOrWc->nTerm-1)?-1:ii); 4728 int r; 4729 r = sqlite3ExprCodeGetColumn(pParse, pTabItem->pTab, -1, iCur, 4730 regRowid, 0); 4731 sqlite3VdbeAddOp4Int(v, OP_RowSetTest, regRowset, 4732 sqlite3VdbeCurrentAddr(v)+2, r, iSet); 4733 } 4734 sqlite3VdbeAddOp2(v, OP_Gosub, regReturn, iLoopBody); 4735 4736 /* The pSubWInfo->untestedTerms flag means that this OR term 4737 ** contained one or more AND term from a notReady table. The 4738 ** terms from the notReady table could not be tested and will 4739 ** need to be tested later. 4740 */ 4741 if( pSubWInfo->untestedTerms ) untestedTerms = 1; 4742 4743 /* If all of the OR-connected terms are optimized using the same 4744 ** index, and the index is opened using the same cursor number 4745 ** by each call to sqlite3WhereBegin() made by this loop, it may 4746 ** be possible to use that index as a covering index. 4747 ** 4748 ** If the call to sqlite3WhereBegin() above resulted in a scan that 4749 ** uses an index, and this is either the first OR-connected term 4750 ** processed or the index is the same as that used by all previous 4751 ** terms, set pCov to the candidate covering index. Otherwise, set 4752 ** pCov to NULL to indicate that no candidate covering index will 4753 ** be available. 4754 */ 4755 pLvl = &pSubWInfo->a[0]; 4756 if( (pLvl->plan.wsFlags & WHERE_INDEXED)!=0 4757 && (pLvl->plan.wsFlags & WHERE_TEMP_INDEX)==0 4758 && (ii==0 || pLvl->plan.u.pIdx==pCov) 4759 ){ 4760 assert( pLvl->iIdxCur==iCovCur ); 4761 pCov = pLvl->plan.u.pIdx; 4762 }else{ 4763 pCov = 0; 4764 } 4765 4766 /* Finish the loop through table entries that match term pOrTerm. */ 4767 sqlite3WhereEnd(pSubWInfo); 4768 } 4769 } 4770 } 4771 pLevel->u.pCovidx = pCov; 4772 if( pCov ) pLevel->iIdxCur = iCovCur; 4773 if( pAndExpr ){ 4774 pAndExpr->pLeft = 0; 4775 sqlite3ExprDelete(pParse->db, pAndExpr); 4776 } 4777 sqlite3VdbeChangeP1(v, iRetInit, sqlite3VdbeCurrentAddr(v)); 4778 sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrBrk); 4779 sqlite3VdbeResolveLabel(v, iLoopBody); 4780 4781 if( pWInfo->nLevel>1 ) sqlite3StackFree(pParse->db, pOrTab); 4782 if( !untestedTerms ) disableTerm(pLevel, pTerm); 4783 }else 4784 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */ 4785 4786 { 4787 /* Case 5: There is no usable index. We must do a complete 4788 ** scan of the entire table. 4789 */ 4790 static const u8 aStep[] = { OP_Next, OP_Prev }; 4791 static const u8 aStart[] = { OP_Rewind, OP_Last }; 4792 assert( bRev==0 || bRev==1 ); 4793 assert( omitTable==0 ); 4794 pLevel->op = aStep[bRev]; 4795 pLevel->p1 = iCur; 4796 pLevel->p2 = 1 + sqlite3VdbeAddOp2(v, aStart[bRev], iCur, addrBrk); 4797 pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP; 4798 } 4799 notReady &= ~getMask(pWC->pMaskSet, iCur); 4800 4801 /* Insert code to test every subexpression that can be completely 4802 ** computed using the current set of tables. 4803 ** 4804 ** IMPLEMENTATION-OF: R-49525-50935 Terms that cannot be satisfied through 4805 ** the use of indices become tests that are evaluated against each row of 4806 ** the relevant input tables. 4807 */ 4808 for(pTerm=pWC->a, j=pWC->nTerm; j>0; j--, pTerm++){ 4809 Expr *pE; 4810 testcase( pTerm->wtFlags & TERM_VIRTUAL ); /* IMP: R-30575-11662 */ 4811 testcase( pTerm->wtFlags & TERM_CODED ); 4812 if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue; 4813 if( (pTerm->prereqAll & notReady)!=0 ){ 4814 testcase( pWInfo->untestedTerms==0 4815 && (pWInfo->wctrlFlags & WHERE_ONETABLE_ONLY)!=0 ); 4816 pWInfo->untestedTerms = 1; 4817 continue; 4818 } 4819 pE = pTerm->pExpr; 4820 assert( pE!=0 ); 4821 if( pLevel->iLeftJoin && !ExprHasProperty(pE, EP_FromJoin) ){ 4822 continue; 4823 } 4824 sqlite3ExprIfFalse(pParse, pE, addrCont, SQLITE_JUMPIFNULL); 4825 pTerm->wtFlags |= TERM_CODED; 4826 } 4827 4828 /* For a LEFT OUTER JOIN, generate code that will record the fact that 4829 ** at least one row of the right table has matched the left table. 4830 */ 4831 if( pLevel->iLeftJoin ){ 4832 pLevel->addrFirst = sqlite3VdbeCurrentAddr(v); 4833 sqlite3VdbeAddOp2(v, OP_Integer, 1, pLevel->iLeftJoin); 4834 VdbeComment((v, "record LEFT JOIN hit")); 4835 sqlite3ExprCacheClear(pParse); 4836 for(pTerm=pWC->a, j=0; j<pWC->nTerm; j++, pTerm++){ 4837 testcase( pTerm->wtFlags & TERM_VIRTUAL ); /* IMP: R-30575-11662 */ 4838 testcase( pTerm->wtFlags & TERM_CODED ); 4839 if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue; 4840 if( (pTerm->prereqAll & notReady)!=0 ){ 4841 assert( pWInfo->untestedTerms ); 4842 continue; 4843 } 4844 assert( pTerm->pExpr ); 4845 sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL); 4846 pTerm->wtFlags |= TERM_CODED; 4847 } 4848 } 4849 sqlite3ReleaseTempReg(pParse, iReleaseReg); 4850 4851 return notReady; 4852 } 4853 4854 #if defined(SQLITE_TEST) 4855 /* 4856 ** The following variable holds a text description of query plan generated 4857 ** by the most recent call to sqlite3WhereBegin(). Each call to WhereBegin 4858 ** overwrites the previous. This information is used for testing and 4859 ** analysis only. 4860 */ 4861 char sqlite3_query_plan[BMS*2*40]; /* Text of the join */ 4862 static int nQPlan = 0; /* Next free slow in _query_plan[] */ 4863 4864 #endif /* SQLITE_TEST */ 4865 4866 4867 /* 4868 ** Free a WhereInfo structure 4869 */ 4870 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 4871 if( ALWAYS(pWInfo) ){ 4872 int i; 4873 for(i=0; i<pWInfo->nLevel; i++){ 4874 sqlite3_index_info *pInfo = pWInfo->a[i].pIdxInfo; 4875 if( pInfo ){ 4876 /* assert( pInfo->needToFreeIdxStr==0 || db->mallocFailed ); */ 4877 if( pInfo->needToFreeIdxStr ){ 4878 sqlite3_free(pInfo->idxStr); 4879 } 4880 sqlite3DbFree(db, pInfo); 4881 } 4882 if( pWInfo->a[i].plan.wsFlags & WHERE_TEMP_INDEX ){ 4883 Index *pIdx = pWInfo->a[i].plan.u.pIdx; 4884 if( pIdx ){ 4885 sqlite3DbFree(db, pIdx->zColAff); 4886 sqlite3DbFree(db, pIdx); 4887 } 4888 } 4889 } 4890 whereClauseClear(pWInfo->pWC); 4891 sqlite3DbFree(db, pWInfo); 4892 } 4893 } 4894 4895 4896 /* 4897 ** Generate the beginning of the loop used for WHERE clause processing. 4898 ** The return value is a pointer to an opaque structure that contains 4899 ** information needed to terminate the loop. Later, the calling routine 4900 ** should invoke sqlite3WhereEnd() with the return value of this function 4901 ** in order to complete the WHERE clause processing. 4902 ** 4903 ** If an error occurs, this routine returns NULL. 4904 ** 4905 ** The basic idea is to do a nested loop, one loop for each table in 4906 ** the FROM clause of a select. (INSERT and UPDATE statements are the 4907 ** same as a SELECT with only a single table in the FROM clause.) For 4908 ** example, if the SQL is this: 4909 ** 4910 ** SELECT * FROM t1, t2, t3 WHERE ...; 4911 ** 4912 ** Then the code generated is conceptually like the following: 4913 ** 4914 ** foreach row1 in t1 do \ Code generated 4915 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 4916 ** foreach row3 in t3 do / 4917 ** ... 4918 ** end \ Code generated 4919 ** end |-- by sqlite3WhereEnd() 4920 ** end / 4921 ** 4922 ** Note that the loops might not be nested in the order in which they 4923 ** appear in the FROM clause if a different order is better able to make 4924 ** use of indices. Note also that when the IN operator appears in 4925 ** the WHERE clause, it might result in additional nested loops for 4926 ** scanning through all values on the right-hand side of the IN. 4927 ** 4928 ** There are Btree cursors associated with each table. t1 uses cursor 4929 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 4930 ** And so forth. This routine generates code to open those VDBE cursors 4931 ** and sqlite3WhereEnd() generates the code to close them. 4932 ** 4933 ** The code that sqlite3WhereBegin() generates leaves the cursors named 4934 ** in pTabList pointing at their appropriate entries. The [...] code 4935 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 4936 ** data from the various tables of the loop. 4937 ** 4938 ** If the WHERE clause is empty, the foreach loops must each scan their 4939 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 4940 ** the tables have indices and there are terms in the WHERE clause that 4941 ** refer to those indices, a complete table scan can be avoided and the 4942 ** code will run much faster. Most of the work of this routine is checking 4943 ** to see if there are indices that can be used to speed up the loop. 4944 ** 4945 ** Terms of the WHERE clause are also used to limit which rows actually 4946 ** make it to the "..." in the middle of the loop. After each "foreach", 4947 ** terms of the WHERE clause that use only terms in that loop and outer 4948 ** loops are evaluated and if false a jump is made around all subsequent 4949 ** inner loops (or around the "..." if the test occurs within the inner- 4950 ** most loop) 4951 ** 4952 ** OUTER JOINS 4953 ** 4954 ** An outer join of tables t1 and t2 is conceptally coded as follows: 4955 ** 4956 ** foreach row1 in t1 do 4957 ** flag = 0 4958 ** foreach row2 in t2 do 4959 ** start: 4960 ** ... 4961 ** flag = 1 4962 ** end 4963 ** if flag==0 then 4964 ** move the row2 cursor to a null row 4965 ** goto start 4966 ** fi 4967 ** end 4968 ** 4969 ** ORDER BY CLAUSE PROCESSING 4970 ** 4971 ** pOrderBy is a pointer to the ORDER BY clause of a SELECT statement, 4972 ** if there is one. If there is no ORDER BY clause or if this routine 4973 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 4974 ** 4975 ** If an index can be used so that the natural output order of the table 4976 ** scan is correct for the ORDER BY clause, then that index is used and 4977 ** the returned WhereInfo.nOBSat field is set to pOrderBy->nExpr. This 4978 ** is an optimization that prevents an unnecessary sort of the result set 4979 ** if an index appropriate for the ORDER BY clause already exists. 4980 ** 4981 ** If the where clause loops cannot be arranged to provide the correct 4982 ** output order, then WhereInfo.nOBSat is 0. 4983 */ 4984 WhereInfo *sqlite3WhereBegin( 4985 Parse *pParse, /* The parser context */ 4986 SrcList *pTabList, /* A list of all tables to be scanned */ 4987 Expr *pWhere, /* The WHERE clause */ 4988 ExprList *pOrderBy, /* An ORDER BY clause, or NULL */ 4989 ExprList *pDistinct, /* The select-list for DISTINCT queries - or NULL */ 4990 u16 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */ 4991 int iIdxCur /* If WHERE_ONETABLE_ONLY is set, index cursor number */ 4992 ){ 4993 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4994 int nTabList; /* Number of elements in pTabList */ 4995 WhereInfo *pWInfo; /* Will become the return value of this function */ 4996 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4997 Bitmask notReady; /* Cursors that are not yet positioned */ 4998 WhereBestIdx sWBI; /* Best index search context */ 4999 WhereMaskSet *pMaskSet; /* The expression mask set */ 5000 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 5001 int iFrom; /* First unused FROM clause element */ 5002 int andFlags; /* AND-ed combination of all pWC->a[].wtFlags */ 5003 int ii; /* Loop counter */ 5004 sqlite3 *db; /* Database connection */ 5005 5006 5007 /* Variable initialization */ 5008 memset(&sWBI, 0, sizeof(sWBI)); 5009 sWBI.pParse = pParse; 5010 5011 /* The number of tables in the FROM clause is limited by the number of 5012 ** bits in a Bitmask 5013 */ 5014 testcase( pTabList->nSrc==BMS ); 5015 if( pTabList->nSrc>BMS ){ 5016 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 5017 return 0; 5018 } 5019 5020 /* This function normally generates a nested loop for all tables in 5021 ** pTabList. But if the WHERE_ONETABLE_ONLY flag is set, then we should 5022 ** only generate code for the first table in pTabList and assume that 5023 ** any cursors associated with subsequent tables are uninitialized. 5024 */ 5025 nTabList = (wctrlFlags & WHERE_ONETABLE_ONLY) ? 1 : pTabList->nSrc; 5026 5027 /* Allocate and initialize the WhereInfo structure that will become the 5028 ** return value. A single allocation is used to store the WhereInfo 5029 ** struct, the contents of WhereInfo.a[], the WhereClause structure 5030 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 5031 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 5032 ** some architectures. Hence the ROUND8() below. 5033 */ 5034 db = pParse->db; 5035 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 5036 pWInfo = sqlite3DbMallocZero(db, 5037 nByteWInfo + 5038 sizeof(WhereClause) + 5039 sizeof(WhereMaskSet) 5040 ); 5041 if( db->mallocFailed ){ 5042 sqlite3DbFree(db, pWInfo); 5043 pWInfo = 0; 5044 goto whereBeginError; 5045 } 5046 pWInfo->nLevel = nTabList; 5047 pWInfo->pParse = pParse; 5048 pWInfo->pTabList = pTabList; 5049 pWInfo->iBreak = sqlite3VdbeMakeLabel(v); 5050 pWInfo->pWC = sWBI.pWC = (WhereClause *)&((u8 *)pWInfo)[nByteWInfo]; 5051 pWInfo->wctrlFlags = wctrlFlags; 5052 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 5053 pMaskSet = (WhereMaskSet*)&sWBI.pWC[1]; 5054 sWBI.aLevel = pWInfo->a; 5055 5056 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 5057 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 5058 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ) pDistinct = 0; 5059 5060 /* Split the WHERE clause into separate subexpressions where each 5061 ** subexpression is separated by an AND operator. 5062 */ 5063 initMaskSet(pMaskSet); 5064 whereClauseInit(sWBI.pWC, pParse, pMaskSet, wctrlFlags); 5065 sqlite3ExprCodeConstants(pParse, pWhere); 5066 whereSplit(sWBI.pWC, pWhere, TK_AND); /* IMP: R-15842-53296 */ 5067 5068 /* Special case: a WHERE clause that is constant. Evaluate the 5069 ** expression and either jump over all of the code or fall thru. 5070 */ 5071 if( pWhere && (nTabList==0 || sqlite3ExprIsConstantNotJoin(pWhere)) ){ 5072 sqlite3ExprIfFalse(pParse, pWhere, pWInfo->iBreak, SQLITE_JUMPIFNULL); 5073 pWhere = 0; 5074 } 5075 5076 /* Assign a bit from the bitmask to every term in the FROM clause. 5077 ** 5078 ** When assigning bitmask values to FROM clause cursors, it must be 5079 ** the case that if X is the bitmask for the N-th FROM clause term then 5080 ** the bitmask for all FROM clause terms to the left of the N-th term 5081 ** is (X-1). An expression from the ON clause of a LEFT JOIN can use 5082 ** its Expr.iRightJoinTable value to find the bitmask of the right table 5083 ** of the join. Subtracting one from the right table bitmask gives a 5084 ** bitmask for all tables to the left of the join. Knowing the bitmask 5085 ** for all tables to the left of a left join is important. Ticket #3015. 5086 ** 5087 ** Note that bitmasks are created for all pTabList->nSrc tables in 5088 ** pTabList, not just the first nTabList tables. nTabList is normally 5089 ** equal to pTabList->nSrc but might be shortened to 1 if the 5090 ** WHERE_ONETABLE_ONLY flag is set. 5091 */ 5092 for(ii=0; ii<pTabList->nSrc; ii++){ 5093 createMask(pMaskSet, pTabList->a[ii].iCursor); 5094 } 5095 #ifndef NDEBUG 5096 { 5097 Bitmask toTheLeft = 0; 5098 for(ii=0; ii<pTabList->nSrc; ii++){ 5099 Bitmask m = getMask(pMaskSet, pTabList->a[ii].iCursor); 5100 assert( (m-1)==toTheLeft ); 5101 toTheLeft |= m; 5102 } 5103 } 5104 #endif 5105 5106 /* Analyze all of the subexpressions. Note that exprAnalyze() might 5107 ** add new virtual terms onto the end of the WHERE clause. We do not 5108 ** want to analyze these virtual terms, so start analyzing at the end 5109 ** and work forward so that the added virtual terms are never processed. 5110 */ 5111 exprAnalyzeAll(pTabList, sWBI.pWC); 5112 if( db->mallocFailed ){ 5113 goto whereBeginError; 5114 } 5115 5116 /* Check if the DISTINCT qualifier, if there is one, is redundant. 5117 ** If it is, then set pDistinct to NULL and WhereInfo.eDistinct to 5118 ** WHERE_DISTINCT_UNIQUE to tell the caller to ignore the DISTINCT. 5119 */ 5120 if( pDistinct && isDistinctRedundant(pParse, pTabList, sWBI.pWC, pDistinct) ){ 5121 pDistinct = 0; 5122 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 5123 } 5124 5125 /* Chose the best index to use for each table in the FROM clause. 5126 ** 5127 ** This loop fills in the following fields: 5128 ** 5129 ** pWInfo->a[].pIdx The index to use for this level of the loop. 5130 ** pWInfo->a[].wsFlags WHERE_xxx flags associated with pIdx 5131 ** pWInfo->a[].nEq The number of == and IN constraints 5132 ** pWInfo->a[].iFrom Which term of the FROM clause is being coded 5133 ** pWInfo->a[].iTabCur The VDBE cursor for the database table 5134 ** pWInfo->a[].iIdxCur The VDBE cursor for the index 5135 ** pWInfo->a[].pTerm When wsFlags==WO_OR, the OR-clause term 5136 ** 5137 ** This loop also figures out the nesting order of tables in the FROM 5138 ** clause. 5139 */ 5140 sWBI.notValid = ~(Bitmask)0; 5141 sWBI.pOrderBy = pOrderBy; 5142 sWBI.n = nTabList; 5143 sWBI.pDistinct = pDistinct; 5144 andFlags = ~0; 5145 WHERETRACE(("*** Optimizer Start ***\n")); 5146 for(sWBI.i=iFrom=0, pLevel=pWInfo->a; sWBI.i<nTabList; sWBI.i++, pLevel++){ 5147 WhereCost bestPlan; /* Most efficient plan seen so far */ 5148 Index *pIdx; /* Index for FROM table at pTabItem */ 5149 int j; /* For looping over FROM tables */ 5150 int bestJ = -1; /* The value of j */ 5151 Bitmask m; /* Bitmask value for j or bestJ */ 5152 int isOptimal; /* Iterator for optimal/non-optimal search */ 5153 int ckOptimal; /* Do the optimal scan check */ 5154 int nUnconstrained; /* Number tables without INDEXED BY */ 5155 Bitmask notIndexed; /* Mask of tables that cannot use an index */ 5156 5157 memset(&bestPlan, 0, sizeof(bestPlan)); 5158 bestPlan.rCost = SQLITE_BIG_DBL; 5159 WHERETRACE(("*** Begin search for loop %d ***\n", sWBI.i)); 5160 5161 /* Loop through the remaining entries in the FROM clause to find the 5162 ** next nested loop. The loop tests all FROM clause entries 5163 ** either once or twice. 5164 ** 5165 ** The first test is always performed if there are two or more entries 5166 ** remaining and never performed if there is only one FROM clause entry 5167 ** to choose from. The first test looks for an "optimal" scan. In 5168 ** this context an optimal scan is one that uses the same strategy 5169 ** for the given FROM clause entry as would be selected if the entry 5170 ** were used as the innermost nested loop. In other words, a table 5171 ** is chosen such that the cost of running that table cannot be reduced 5172 ** by waiting for other tables to run first. This "optimal" test works 5173 ** by first assuming that the FROM clause is on the inner loop and finding 5174 ** its query plan, then checking to see if that query plan uses any 5175 ** other FROM clause terms that are sWBI.notValid. If no notValid terms 5176 ** are used then the "optimal" query plan works. 5177 ** 5178 ** Note that the WhereCost.nRow parameter for an optimal scan might 5179 ** not be as small as it would be if the table really were the innermost 5180 ** join. The nRow value can be reduced by WHERE clause constraints 5181 ** that do not use indices. But this nRow reduction only happens if the 5182 ** table really is the innermost join. 5183 ** 5184 ** The second loop iteration is only performed if no optimal scan 5185 ** strategies were found by the first iteration. This second iteration 5186 ** is used to search for the lowest cost scan overall. 5187 ** 5188 ** Without the optimal scan step (the first iteration) a suboptimal 5189 ** plan might be chosen for queries like this: 5190 ** 5191 ** CREATE TABLE t1(a, b); 5192 ** CREATE TABLE t2(c, d); 5193 ** SELECT * FROM t2, t1 WHERE t2.rowid = t1.a; 5194 ** 5195 ** The best strategy is to iterate through table t1 first. However it 5196 ** is not possible to determine this with a simple greedy algorithm. 5197 ** Since the cost of a linear scan through table t2 is the same 5198 ** as the cost of a linear scan through table t1, a simple greedy 5199 ** algorithm may choose to use t2 for the outer loop, which is a much 5200 ** costlier approach. 5201 */ 5202 nUnconstrained = 0; 5203 notIndexed = 0; 5204 5205 /* The optimal scan check only occurs if there are two or more tables 5206 ** available to be reordered */ 5207 if( iFrom==nTabList-1 ){ 5208 ckOptimal = 0; /* Common case of just one table in the FROM clause */ 5209 }else{ 5210 ckOptimal = -1; 5211 for(j=iFrom, sWBI.pSrc=&pTabList->a[j]; j<nTabList; j++, sWBI.pSrc++){ 5212 m = getMask(pMaskSet, sWBI.pSrc->iCursor); 5213 if( (m & sWBI.notValid)==0 ){ 5214 if( j==iFrom ) iFrom++; 5215 continue; 5216 } 5217 if( j>iFrom && (sWBI.pSrc->jointype & (JT_LEFT|JT_CROSS))!=0 ) break; 5218 if( ++ckOptimal ) break; 5219 if( (sWBI.pSrc->jointype & JT_LEFT)!=0 ) break; 5220 } 5221 } 5222 assert( ckOptimal==0 || ckOptimal==1 ); 5223 5224 for(isOptimal=ckOptimal; isOptimal>=0 && bestJ<0; isOptimal--){ 5225 for(j=iFrom, sWBI.pSrc=&pTabList->a[j]; j<nTabList; j++, sWBI.pSrc++){ 5226 if( j>iFrom && (sWBI.pSrc->jointype & (JT_LEFT|JT_CROSS))!=0 ){ 5227 /* This break and one like it in the ckOptimal computation loop 5228 ** above prevent table reordering across LEFT and CROSS JOINs. 5229 ** The LEFT JOIN case is necessary for correctness. The prohibition 5230 ** against reordering across a CROSS JOIN is an SQLite feature that 5231 ** allows the developer to control table reordering */ 5232 break; 5233 } 5234 m = getMask(pMaskSet, sWBI.pSrc->iCursor); 5235 if( (m & sWBI.notValid)==0 ){ 5236 assert( j>iFrom ); 5237 continue; 5238 } 5239 sWBI.notReady = (isOptimal ? m : sWBI.notValid); 5240 if( sWBI.pSrc->pIndex==0 ) nUnconstrained++; 5241 5242 WHERETRACE((" === trying table %d (%s) with isOptimal=%d ===\n", 5243 j, sWBI.pSrc->pTab->zName, isOptimal)); 5244 assert( sWBI.pSrc->pTab ); 5245 #ifndef SQLITE_OMIT_VIRTUALTABLE 5246 if( IsVirtual(sWBI.pSrc->pTab) ){ 5247 sWBI.ppIdxInfo = &pWInfo->a[j].pIdxInfo; 5248 bestVirtualIndex(&sWBI); 5249 }else 5250 #endif 5251 { 5252 bestBtreeIndex(&sWBI); 5253 } 5254 assert( isOptimal || (sWBI.cost.used&sWBI.notValid)==0 ); 5255 5256 /* If an INDEXED BY clause is present, then the plan must use that 5257 ** index if it uses any index at all */ 5258 assert( sWBI.pSrc->pIndex==0 5259 || (sWBI.cost.plan.wsFlags & WHERE_NOT_FULLSCAN)==0 5260 || sWBI.cost.plan.u.pIdx==sWBI.pSrc->pIndex ); 5261 5262 if( isOptimal && (sWBI.cost.plan.wsFlags & WHERE_NOT_FULLSCAN)==0 ){ 5263 notIndexed |= m; 5264 } 5265 if( isOptimal ){ 5266 pWInfo->a[j].rOptCost = sWBI.cost.rCost; 5267 }else if( ckOptimal ){ 5268 /* If two or more tables have nearly the same outer loop cost, but 5269 ** very different inner loop (optimal) cost, we want to choose 5270 ** for the outer loop that table which benefits the least from 5271 ** being in the inner loop. The following code scales the 5272 ** outer loop cost estimate to accomplish that. */ 5273 WHERETRACE((" scaling cost from %.1f to %.1f\n", 5274 sWBI.cost.rCost, 5275 sWBI.cost.rCost/pWInfo->a[j].rOptCost)); 5276 sWBI.cost.rCost /= pWInfo->a[j].rOptCost; 5277 } 5278 5279 /* Conditions under which this table becomes the best so far: 5280 ** 5281 ** (1) The table must not depend on other tables that have not 5282 ** yet run. (In other words, it must not depend on tables 5283 ** in inner loops.) 5284 ** 5285 ** (2) (This rule was removed on 2012-11-09. The scaling of the 5286 ** cost using the optimal scan cost made this rule obsolete.) 5287 ** 5288 ** (3) All tables have an INDEXED BY clause or this table lacks an 5289 ** INDEXED BY clause or this table uses the specific 5290 ** index specified by its INDEXED BY clause. This rule ensures 5291 ** that a best-so-far is always selected even if an impossible 5292 ** combination of INDEXED BY clauses are given. The error 5293 ** will be detected and relayed back to the application later. 5294 ** The NEVER() comes about because rule (2) above prevents 5295 ** An indexable full-table-scan from reaching rule (3). 5296 ** 5297 ** (4) The plan cost must be lower than prior plans, where "cost" 5298 ** is defined by the compareCost() function above. 5299 */ 5300 if( (sWBI.cost.used&sWBI.notValid)==0 /* (1) */ 5301 && (nUnconstrained==0 || sWBI.pSrc->pIndex==0 /* (3) */ 5302 || NEVER((sWBI.cost.plan.wsFlags & WHERE_NOT_FULLSCAN)!=0)) 5303 && (bestJ<0 || compareCost(&sWBI.cost, &bestPlan)) /* (4) */ 5304 ){ 5305 WHERETRACE((" === table %d (%s) is best so far\n" 5306 " cost=%.1f, nRow=%.1f, nOBSat=%d, wsFlags=%08x\n", 5307 j, sWBI.pSrc->pTab->zName, 5308 sWBI.cost.rCost, sWBI.cost.plan.nRow, 5309 sWBI.cost.plan.nOBSat, sWBI.cost.plan.wsFlags)); 5310 bestPlan = sWBI.cost; 5311 bestJ = j; 5312 } 5313 5314 /* In a join like "w JOIN x LEFT JOIN y JOIN z" make sure that 5315 ** table y (and not table z) is always the next inner loop inside 5316 ** of table x. */ 5317 if( (sWBI.pSrc->jointype & JT_LEFT)!=0 ) break; 5318 } 5319 } 5320 assert( bestJ>=0 ); 5321 assert( sWBI.notValid & getMask(pMaskSet, pTabList->a[bestJ].iCursor) ); 5322 assert( bestJ==iFrom || (pTabList->a[iFrom].jointype & JT_LEFT)==0 ); 5323 testcase( bestJ>iFrom && (pTabList->a[iFrom].jointype & JT_CROSS)!=0 ); 5324 testcase( bestJ>iFrom && bestJ<nTabList-1 5325 && (pTabList->a[bestJ+1].jointype & JT_LEFT)!=0 ); 5326 WHERETRACE(("*** Optimizer selects table %d (%s) for loop %d with:\n" 5327 " cost=%.1f, nRow=%.1f, nOBSat=%d, wsFlags=0x%08x\n", 5328 bestJ, pTabList->a[bestJ].pTab->zName, 5329 pLevel-pWInfo->a, bestPlan.rCost, bestPlan.plan.nRow, 5330 bestPlan.plan.nOBSat, bestPlan.plan.wsFlags)); 5331 if( (bestPlan.plan.wsFlags & WHERE_DISTINCT)!=0 ){ 5332 assert( pWInfo->eDistinct==0 ); 5333 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 5334 } 5335 andFlags &= bestPlan.plan.wsFlags; 5336 pLevel->plan = bestPlan.plan; 5337 pLevel->iTabCur = pTabList->a[bestJ].iCursor; 5338 testcase( bestPlan.plan.wsFlags & WHERE_INDEXED ); 5339 testcase( bestPlan.plan.wsFlags & WHERE_TEMP_INDEX ); 5340 if( bestPlan.plan.wsFlags & (WHERE_INDEXED|WHERE_TEMP_INDEX) ){ 5341 if( (wctrlFlags & WHERE_ONETABLE_ONLY) 5342 && (bestPlan.plan.wsFlags & WHERE_TEMP_INDEX)==0 5343 ){ 5344 pLevel->iIdxCur = iIdxCur; 5345 }else{ 5346 pLevel->iIdxCur = pParse->nTab++; 5347 } 5348 }else{ 5349 pLevel->iIdxCur = -1; 5350 } 5351 sWBI.notValid &= ~getMask(pMaskSet, pTabList->a[bestJ].iCursor); 5352 pLevel->iFrom = (u8)bestJ; 5353 if( bestPlan.plan.nRow>=(double)1 ){ 5354 pParse->nQueryLoop *= bestPlan.plan.nRow; 5355 } 5356 5357 /* Check that if the table scanned by this loop iteration had an 5358 ** INDEXED BY clause attached to it, that the named index is being 5359 ** used for the scan. If not, then query compilation has failed. 5360 ** Return an error. 5361 */ 5362 pIdx = pTabList->a[bestJ].pIndex; 5363 if( pIdx ){ 5364 if( (bestPlan.plan.wsFlags & WHERE_INDEXED)==0 ){ 5365 sqlite3ErrorMsg(pParse, "cannot use index: %s", pIdx->zName); 5366 goto whereBeginError; 5367 }else{ 5368 /* If an INDEXED BY clause is used, the bestIndex() function is 5369 ** guaranteed to find the index specified in the INDEXED BY clause 5370 ** if it find an index at all. */ 5371 assert( bestPlan.plan.u.pIdx==pIdx ); 5372 } 5373 } 5374 } 5375 WHERETRACE(("*** Optimizer Finished ***\n")); 5376 if( pParse->nErr || db->mallocFailed ){ 5377 goto whereBeginError; 5378 } 5379 if( nTabList ){ 5380 pLevel--; 5381 pWInfo->nOBSat = pLevel->plan.nOBSat; 5382 }else{ 5383 pWInfo->nOBSat = 0; 5384 } 5385 5386 /* If the total query only selects a single row, then the ORDER BY 5387 ** clause is irrelevant. 5388 */ 5389 if( (andFlags & WHERE_UNIQUE)!=0 && pOrderBy ){ 5390 assert( nTabList==0 || (pLevel->plan.wsFlags & WHERE_ALL_UNIQUE)!=0 ); 5391 pWInfo->nOBSat = pOrderBy->nExpr; 5392 } 5393 5394 /* If the caller is an UPDATE or DELETE statement that is requesting 5395 ** to use a one-pass algorithm, determine if this is appropriate. 5396 ** The one-pass algorithm only works if the WHERE clause constraints 5397 ** the statement to update a single row. 5398 */ 5399 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 5400 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 && (andFlags & WHERE_UNIQUE)!=0 ){ 5401 pWInfo->okOnePass = 1; 5402 pWInfo->a[0].plan.wsFlags &= ~WHERE_IDX_ONLY; 5403 } 5404 5405 /* Open all tables in the pTabList and any indices selected for 5406 ** searching those tables. 5407 */ 5408 sqlite3CodeVerifySchema(pParse, -1); /* Insert the cookie verifier Goto */ 5409 notReady = ~(Bitmask)0; 5410 pWInfo->nRowOut = (double)1; 5411 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 5412 Table *pTab; /* Table to open */ 5413 int iDb; /* Index of database containing table/index */ 5414 struct SrcList_item *pTabItem; 5415 5416 pTabItem = &pTabList->a[pLevel->iFrom]; 5417 pTab = pTabItem->pTab; 5418 pWInfo->nRowOut *= pLevel->plan.nRow; 5419 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 5420 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 5421 /* Do nothing */ 5422 }else 5423 #ifndef SQLITE_OMIT_VIRTUALTABLE 5424 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 5425 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 5426 int iCur = pTabItem->iCursor; 5427 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 5428 }else if( IsVirtual(pTab) ){ 5429 /* noop */ 5430 }else 5431 #endif 5432 if( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 5433 && (wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 ){ 5434 int op = pWInfo->okOnePass ? OP_OpenWrite : OP_OpenRead; 5435 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 5436 testcase( pTab->nCol==BMS-1 ); 5437 testcase( pTab->nCol==BMS ); 5438 if( !pWInfo->okOnePass && pTab->nCol<BMS ){ 5439 Bitmask b = pTabItem->colUsed; 5440 int n = 0; 5441 for(; b; b=b>>1, n++){} 5442 sqlite3VdbeChangeP4(v, sqlite3VdbeCurrentAddr(v)-1, 5443 SQLITE_INT_TO_PTR(n), P4_INT32); 5444 assert( n<=pTab->nCol ); 5445 } 5446 }else{ 5447 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 5448 } 5449 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 5450 if( (pLevel->plan.wsFlags & WHERE_TEMP_INDEX)!=0 ){ 5451 constructAutomaticIndex(pParse, sWBI.pWC, pTabItem, notReady, pLevel); 5452 }else 5453 #endif 5454 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){ 5455 Index *pIx = pLevel->plan.u.pIdx; 5456 KeyInfo *pKey = sqlite3IndexKeyinfo(pParse, pIx); 5457 int iIndexCur = pLevel->iIdxCur; 5458 assert( pIx->pSchema==pTab->pSchema ); 5459 assert( iIndexCur>=0 ); 5460 sqlite3VdbeAddOp4(v, OP_OpenRead, iIndexCur, pIx->tnum, iDb, 5461 (char*)pKey, P4_KEYINFO_HANDOFF); 5462 VdbeComment((v, "%s", pIx->zName)); 5463 } 5464 sqlite3CodeVerifySchema(pParse, iDb); 5465 notReady &= ~getMask(sWBI.pWC->pMaskSet, pTabItem->iCursor); 5466 } 5467 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 5468 if( db->mallocFailed ) goto whereBeginError; 5469 5470 /* Generate the code to do the search. Each iteration of the for 5471 ** loop below generates code for a single nested loop of the VM 5472 ** program. 5473 */ 5474 notReady = ~(Bitmask)0; 5475 for(ii=0; ii<nTabList; ii++){ 5476 pLevel = &pWInfo->a[ii]; 5477 explainOneScan(pParse, pTabList, pLevel, ii, pLevel->iFrom, wctrlFlags); 5478 notReady = codeOneLoopStart(pWInfo, ii, wctrlFlags, notReady); 5479 pWInfo->iContinue = pLevel->addrCont; 5480 } 5481 5482 #ifdef SQLITE_TEST /* For testing and debugging use only */ 5483 /* Record in the query plan information about the current table 5484 ** and the index used to access it (if any). If the table itself 5485 ** is not used, its name is just '{}'. If no index is used 5486 ** the index is listed as "{}". If the primary key is used the 5487 ** index name is '*'. 5488 */ 5489 for(ii=0; ii<nTabList; ii++){ 5490 char *z; 5491 int n; 5492 int w; 5493 struct SrcList_item *pTabItem; 5494 5495 pLevel = &pWInfo->a[ii]; 5496 w = pLevel->plan.wsFlags; 5497 pTabItem = &pTabList->a[pLevel->iFrom]; 5498 z = pTabItem->zAlias; 5499 if( z==0 ) z = pTabItem->pTab->zName; 5500 n = sqlite3Strlen30(z); 5501 if( n+nQPlan < sizeof(sqlite3_query_plan)-10 ){ 5502 if( (w & WHERE_IDX_ONLY)!=0 && (w & WHERE_COVER_SCAN)==0 ){ 5503 memcpy(&sqlite3_query_plan[nQPlan], "{}", 2); 5504 nQPlan += 2; 5505 }else{ 5506 memcpy(&sqlite3_query_plan[nQPlan], z, n); 5507 nQPlan += n; 5508 } 5509 sqlite3_query_plan[nQPlan++] = ' '; 5510 } 5511 testcase( w & WHERE_ROWID_EQ ); 5512 testcase( w & WHERE_ROWID_RANGE ); 5513 if( w & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){ 5514 memcpy(&sqlite3_query_plan[nQPlan], "* ", 2); 5515 nQPlan += 2; 5516 }else if( (w & WHERE_INDEXED)!=0 && (w & WHERE_COVER_SCAN)==0 ){ 5517 n = sqlite3Strlen30(pLevel->plan.u.pIdx->zName); 5518 if( n+nQPlan < sizeof(sqlite3_query_plan)-2 ){ 5519 memcpy(&sqlite3_query_plan[nQPlan], pLevel->plan.u.pIdx->zName, n); 5520 nQPlan += n; 5521 sqlite3_query_plan[nQPlan++] = ' '; 5522 } 5523 }else{ 5524 memcpy(&sqlite3_query_plan[nQPlan], "{} ", 3); 5525 nQPlan += 3; 5526 } 5527 } 5528 while( nQPlan>0 && sqlite3_query_plan[nQPlan-1]==' ' ){ 5529 sqlite3_query_plan[--nQPlan] = 0; 5530 } 5531 sqlite3_query_plan[nQPlan] = 0; 5532 nQPlan = 0; 5533 #endif /* SQLITE_TEST // Testing and debugging use only */ 5534 5535 /* Record the continuation address in the WhereInfo structure. Then 5536 ** clean up and return. 5537 */ 5538 return pWInfo; 5539 5540 /* Jump here if malloc fails */ 5541 whereBeginError: 5542 if( pWInfo ){ 5543 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5544 whereInfoFree(db, pWInfo); 5545 } 5546 return 0; 5547 } 5548 5549 /* 5550 ** Generate the end of the WHERE loop. See comments on 5551 ** sqlite3WhereBegin() for additional information. 5552 */ 5553 void sqlite3WhereEnd(WhereInfo *pWInfo){ 5554 Parse *pParse = pWInfo->pParse; 5555 Vdbe *v = pParse->pVdbe; 5556 int i; 5557 WhereLevel *pLevel; 5558 SrcList *pTabList = pWInfo->pTabList; 5559 sqlite3 *db = pParse->db; 5560 5561 /* Generate loop termination code. 5562 */ 5563 sqlite3ExprCacheClear(pParse); 5564 for(i=pWInfo->nLevel-1; i>=0; i--){ 5565 pLevel = &pWInfo->a[i]; 5566 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5567 if( pLevel->op!=OP_Noop ){ 5568 sqlite3VdbeAddOp2(v, pLevel->op, pLevel->p1, pLevel->p2); 5569 sqlite3VdbeChangeP5(v, pLevel->p5); 5570 } 5571 if( pLevel->plan.wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 5572 struct InLoop *pIn; 5573 int j; 5574 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 5575 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 5576 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5577 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 5578 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 5579 } 5580 sqlite3DbFree(db, pLevel->u.in.aInLoop); 5581 } 5582 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 5583 if( pLevel->iLeftJoin ){ 5584 int addr; 5585 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); 5586 assert( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 5587 || (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ); 5588 if( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 ){ 5589 sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor); 5590 } 5591 if( pLevel->iIdxCur>=0 ){ 5592 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 5593 } 5594 if( pLevel->op==OP_Return ){ 5595 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 5596 }else{ 5597 sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrFirst); 5598 } 5599 sqlite3VdbeJumpHere(v, addr); 5600 } 5601 } 5602 5603 /* The "break" point is here, just past the end of the outer loop. 5604 ** Set it. 5605 */ 5606 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5607 5608 /* Close all of the cursors that were opened by sqlite3WhereBegin. 5609 */ 5610 assert( pWInfo->nLevel==1 || pWInfo->nLevel==pTabList->nSrc ); 5611 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5612 Index *pIdx = 0; 5613 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; 5614 Table *pTab = pTabItem->pTab; 5615 assert( pTab!=0 ); 5616 if( (pTab->tabFlags & TF_Ephemeral)==0 5617 && pTab->pSelect==0 5618 && (pWInfo->wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 5619 ){ 5620 int ws = pLevel->plan.wsFlags; 5621 if( !pWInfo->okOnePass && (ws & WHERE_IDX_ONLY)==0 ){ 5622 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); 5623 } 5624 if( (ws & WHERE_INDEXED)!=0 && (ws & WHERE_TEMP_INDEX)==0 ){ 5625 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); 5626 } 5627 } 5628 5629 /* If this scan uses an index, make code substitutions to read data 5630 ** from the index in preference to the table. Sometimes, this means 5631 ** the table need never be read from. This is a performance boost, 5632 ** as the vdbe level waits until the table is read before actually 5633 ** seeking the table cursor to the record corresponding to the current 5634 ** position in the index. 5635 ** 5636 ** Calls to the code generator in between sqlite3WhereBegin and 5637 ** sqlite3WhereEnd will have created code that references the table 5638 ** directly. This loop scans all that code looking for opcodes 5639 ** that reference the table and converts them into opcodes that 5640 ** reference the index. 5641 */ 5642 if( pLevel->plan.wsFlags & WHERE_INDEXED ){ 5643 pIdx = pLevel->plan.u.pIdx; 5644 }else if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){ 5645 pIdx = pLevel->u.pCovidx; 5646 } 5647 if( pIdx && !db->mallocFailed){ 5648 int k, j, last; 5649 VdbeOp *pOp; 5650 5651 pOp = sqlite3VdbeGetOp(v, pWInfo->iTop); 5652 last = sqlite3VdbeCurrentAddr(v); 5653 for(k=pWInfo->iTop; k<last; k++, pOp++){ 5654 if( pOp->p1!=pLevel->iTabCur ) continue; 5655 if( pOp->opcode==OP_Column ){ 5656 for(j=0; j<pIdx->nColumn; j++){ 5657 if( pOp->p2==pIdx->aiColumn[j] ){ 5658 pOp->p2 = j; 5659 pOp->p1 = pLevel->iIdxCur; 5660 break; 5661 } 5662 } 5663 assert( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 5664 || j<pIdx->nColumn ); 5665 }else if( pOp->opcode==OP_Rowid ){ 5666 pOp->p1 = pLevel->iIdxCur; 5667 pOp->opcode = OP_IdxRowid; 5668 } 5669 } 5670 } 5671 } 5672 5673 /* Final cleanup 5674 */ 5675 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5676 whereInfoFree(db, pWInfo); 5677 return; 5678 } 5679