blob: a7244193245111fb9cd3e618ac747a1886b2e556 [file] [log] [blame]
drh75897232000-05-29 14:26:00 +00001/*
drhb19a2bc2001-09-16 00:13:26 +00002** 2001 September 15
drh75897232000-05-29 14:26:00 +00003**
drhb19a2bc2001-09-16 00:13:26 +00004** The author disclaims copyright to this source code. In place of
5** a legal notice, here is a blessing:
drh75897232000-05-29 14:26:00 +00006**
drhb19a2bc2001-09-16 00:13:26 +00007** 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.
drh75897232000-05-29 14:26:00 +000010**
11*************************************************************************
12** This module contains C code that generates VDBE code used to process
drh909626d2008-05-30 14:58:37 +000013** the WHERE clause of SQL statements. This module is responsible for
drh51669862004-12-18 18:40:26 +000014** 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".
drh75897232000-05-29 14:26:00 +000018*/
19#include "sqliteInt.h"
drhe54df422013-11-12 18:37:25 +000020#include "whereInt.h"
drh51147ba2005-07-23 22:59:55 +000021
drhefc88d02017-12-22 00:52:50 +000022/*
23** Extra information appended to the end of sqlite3_index_info but not
24** visible to the xBestIndex function, at least not directly. The
25** sqlite3_vtab_collation() interface knows how to reach it, however.
26**
27** This object is not an API and can be changed from one release to the
28** next. As long as allocateIndexInfo() and sqlite3_vtab_collation()
29** agree on the structure, all will be well.
30*/
31typedef struct HiddenIndexInfo HiddenIndexInfo;
32struct HiddenIndexInfo {
drh82801a52022-01-20 17:10:59 +000033 WhereClause *pWC; /* The Where clause being analyzed */
34 Parse *pParse; /* The parsing context */
drhec778d22022-01-22 00:18:01 +000035 int eDistinct; /* Value to return from sqlite3_vtab_distinct() */
drh0fe7e7d2022-02-01 14:58:29 +000036 u32 mIn; /* Mask of terms that are <col> IN (...) */
37 u32 mHandleIn; /* Terms that vtab will handle as <col> IN (...) */
drh82801a52022-01-20 17:10:59 +000038 sqlite3_value *aRhs[1]; /* RHS values for constraints. MUST BE LAST
39 ** because extra space is allocated to hold up
40 ** to nTerm such values */
drhefc88d02017-12-22 00:52:50 +000041};
42
drh6f82e852015-06-06 20:12:09 +000043/* Forward declaration of methods */
44static int whereLoopResize(sqlite3*, WhereLoop*, int);
45
drh51147ba2005-07-23 22:59:55 +000046/*
drh6f328482013-06-05 23:39:34 +000047** Return the estimated number of output rows from a WHERE clause
48*/
drhc3489bb2016-02-25 16:04:59 +000049LogEst sqlite3WhereOutputRowCount(WhereInfo *pWInfo){
50 return pWInfo->nRowOut;
drh6f328482013-06-05 23:39:34 +000051}
52
53/*
54** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this
55** WHERE clause returns outputs for DISTINCT processing.
56*/
57int sqlite3WhereIsDistinct(WhereInfo *pWInfo){
58 return pWInfo->eDistinct;
59}
60
61/*
drhc37b7682020-07-14 15:30:35 +000062** Return the number of ORDER BY terms that are satisfied by the
63** WHERE clause. A return of 0 means that the output must be
64** completely sorted. A return equal to the number of ORDER BY
65** terms means that no sorting is needed at all. A return that
66** is positive but less than the number of ORDER BY terms means that
67** block sorting is required.
drh6f328482013-06-05 23:39:34 +000068*/
69int sqlite3WhereIsOrdered(WhereInfo *pWInfo){
drhddba0c22014-03-18 20:33:42 +000070 return pWInfo->nOBSat;
drh6f328482013-06-05 23:39:34 +000071}
72
73/*
drh6ee5a7b2018-09-08 20:09:46 +000074** In the ORDER BY LIMIT optimization, if the inner-most loop is known
75** to emit rows in increasing order, and if the last row emitted by the
76** inner-most loop did not fit within the sorter, then we can skip all
77** subsequent rows for the current iteration of the inner loop (because they
78** will not fit in the sorter either) and continue with the second inner
79** loop - the loop immediately outside the inner-most.
drha536df42016-05-19 22:13:37 +000080**
drh6ee5a7b2018-09-08 20:09:46 +000081** When a row does not fit in the sorter (because the sorter already
82** holds LIMIT+OFFSET rows that are smaller), then a jump is made to the
83** label returned by this function.
84**
85** If the ORDER BY LIMIT optimization applies, the jump destination should
86** be the continuation for the second-inner-most loop. If the ORDER BY
87** LIMIT optimization does not apply, then the jump destination should
88** be the continuation for the inner-most loop.
89**
90** It is always safe for this routine to return the continuation of the
91** inner-most loop, in the sense that a correct answer will result.
92** Returning the continuation the second inner loop is an optimization
93** that might make the code run a little faster, but should not change
94** the final answer.
drha536df42016-05-19 22:13:37 +000095*/
drh6ee5a7b2018-09-08 20:09:46 +000096int sqlite3WhereOrderByLimitOptLabel(WhereInfo *pWInfo){
97 WhereLevel *pInner;
98 if( !pWInfo->bOrderedInnerLoop ){
99 /* The ORDER BY LIMIT optimization does not apply. Jump to the
100 ** continuation of the inner-most loop. */
101 return pWInfo->iContinue;
102 }
103 pInner = &pWInfo->a[pWInfo->nLevel-1];
drhf7ded142018-09-08 20:29:04 +0000104 assert( pInner->addrNxt!=0 );
105 return pInner->addrNxt;
drha536df42016-05-19 22:13:37 +0000106}
107
108/*
drhd1930572021-01-13 15:23:17 +0000109** While generating code for the min/max optimization, after handling
110** the aggregate-step call to min() or max(), check to see if any
111** additional looping is required. If the output order is such that
112** we are certain that the correct answer has already been found, then
113** code an OP_Goto to by pass subsequent processing.
114**
115** Any extra OP_Goto that is coded here is an optimization. The
116** correct answer should be obtained regardless. This OP_Goto just
117** makes the answer appear faster.
118*/
119void sqlite3WhereMinMaxOptEarlyOut(Vdbe *v, WhereInfo *pWInfo){
120 WhereLevel *pInner;
drh5870dc82021-01-14 00:53:14 +0000121 int i;
drhd1930572021-01-13 15:23:17 +0000122 if( !pWInfo->bOrderedInnerLoop ) return;
123 if( pWInfo->nOBSat==0 ) return;
drh5870dc82021-01-14 00:53:14 +0000124 for(i=pWInfo->nLevel-1; i>=0; i--){
125 pInner = &pWInfo->a[i];
126 if( (pInner->pWLoop->wsFlags & WHERE_COLUMN_IN)!=0 ){
127 sqlite3VdbeGoto(v, pInner->addrNxt);
128 return;
129 }
drhd1930572021-01-13 15:23:17 +0000130 }
drh5870dc82021-01-14 00:53:14 +0000131 sqlite3VdbeGoto(v, pWInfo->iBreak);
drhd1930572021-01-13 15:23:17 +0000132}
133
134/*
drh6f328482013-06-05 23:39:34 +0000135** Return the VDBE address or label to jump to in order to continue
136** immediately with the next row of a WHERE clause.
137*/
138int sqlite3WhereContinueLabel(WhereInfo *pWInfo){
drha22a75e2014-03-21 18:16:23 +0000139 assert( pWInfo->iContinue!=0 );
drh6f328482013-06-05 23:39:34 +0000140 return pWInfo->iContinue;
141}
142
143/*
144** Return the VDBE address or label to jump to in order to break
145** out of a WHERE loop.
146*/
147int sqlite3WhereBreakLabel(WhereInfo *pWInfo){
148 return pWInfo->iBreak;
149}
150
151/*
drhb0264ee2015-09-14 14:45:50 +0000152** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to
drhbe3da242019-12-29 00:52:41 +0000153** operate directly on the rowids returned by a WHERE clause. Return
drhb0264ee2015-09-14 14:45:50 +0000154** ONEPASS_SINGLE (1) if the statement can operation directly because only
155** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass
156** optimization can be used on multiple
drhfc8d4f92013-11-08 15:19:46 +0000157**
158** If the ONEPASS optimization is used (if this routine returns true)
159** then also write the indices of open cursors used by ONEPASS
160** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data
161** table and iaCur[1] gets the cursor used by an auxiliary index.
162** Either value may be -1, indicating that cursor is not used.
163** Any cursors returned will have been opened for writing.
164**
165** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is
166** unable to use the ONEPASS optimization.
drh6f328482013-06-05 23:39:34 +0000167*/
drhfc8d4f92013-11-08 15:19:46 +0000168int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){
169 memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2);
drha7228212015-09-28 17:05:22 +0000170#ifdef WHERETRACE_ENABLED
171 if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){
172 sqlite3DebugPrintf("%s cursors: %d %d\n",
173 pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI",
174 aiCur[0], aiCur[1]);
175 }
176#endif
drhb0264ee2015-09-14 14:45:50 +0000177 return pWInfo->eOnePass;
drh6f328482013-06-05 23:39:34 +0000178}
179
180/*
drhbe3da242019-12-29 00:52:41 +0000181** Return TRUE if the WHERE loop uses the OP_DeferredSeek opcode to move
182** the data cursor to the row selected by the index cursor.
183*/
184int sqlite3WhereUsesDeferredSeek(WhereInfo *pWInfo){
185 return pWInfo->bDeferredSeek;
186}
187
188/*
drhaa32e3c2013-07-16 21:31:23 +0000189** Move the content of pSrc into pDest
190*/
191static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){
192 pDest->n = pSrc->n;
193 memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0]));
194}
195
196/*
197** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet.
198**
199** The new entry might overwrite an existing entry, or it might be
200** appended, or it might be discarded. Do whatever is the right thing
201** so that pSet keeps the N_OR_COST best entries seen so far.
202*/
203static int whereOrInsert(
204 WhereOrSet *pSet, /* The WhereOrSet to be updated */
205 Bitmask prereq, /* Prerequisites of the new entry */
drhbf539c42013-10-05 18:16:02 +0000206 LogEst rRun, /* Run-cost of the new entry */
207 LogEst nOut /* Number of outputs for the new entry */
drhaa32e3c2013-07-16 21:31:23 +0000208){
209 u16 i;
210 WhereOrCost *p;
211 for(i=pSet->n, p=pSet->a; i>0; i--, p++){
212 if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){
213 goto whereOrInsert_done;
214 }
215 if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){
216 return 0;
217 }
218 }
219 if( pSet->n<N_OR_COST ){
220 p = &pSet->a[pSet->n++];
221 p->nOut = nOut;
222 }else{
223 p = pSet->a;
224 for(i=1; i<pSet->n; i++){
225 if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i;
226 }
227 if( p->rRun<=rRun ) return 0;
228 }
229whereOrInsert_done:
230 p->prereq = prereq;
231 p->rRun = rRun;
232 if( p->nOut>nOut ) p->nOut = nOut;
233 return 1;
234}
235
236/*
drh1398ad32005-01-19 23:24:50 +0000237** Return the bitmask for the given cursor number. Return 0 if
238** iCursor is not in the set.
drh6a3ea0e2003-05-02 14:32:12 +0000239*/
drh6f82e852015-06-06 20:12:09 +0000240Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){
drh6a3ea0e2003-05-02 14:32:12 +0000241 int i;
drhfcd71b62011-04-05 22:08:24 +0000242 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 );
drhdae2a102021-12-02 04:00:45 +0000243 assert( pMaskSet->n>0 || pMaskSet->ix[0]<0 );
244 assert( iCursor>=-1 );
drh844a89b2021-12-02 12:34:05 +0000245 if( pMaskSet->ix[0]==iCursor ){
246 return 1;
247 }
248 for(i=1; i<pMaskSet->n; i++){
drh51669862004-12-18 18:40:26 +0000249 if( pMaskSet->ix[i]==iCursor ){
drh7699d1c2013-06-04 12:42:29 +0000250 return MASKBIT(i);
drh51669862004-12-18 18:40:26 +0000251 }
drh6a3ea0e2003-05-02 14:32:12 +0000252 }
drh6a3ea0e2003-05-02 14:32:12 +0000253 return 0;
254}
255
drhf8bdcfa2022-04-09 03:06:01 +0000256/* Allocate memory that is automatically freed when pWInfo is freed.
257*/
258void *sqlite3WhereMalloc(WhereInfo *pWInfo, u64 nByte){
259 WhereMemBlock *pBlock;
260 pBlock = sqlite3DbMallocRawNN(pWInfo->pParse->db, nByte+sizeof(*pBlock));
261 if( pBlock ){
262 pBlock->pNext = pWInfo->pMemToFree;
263 pBlock->sz = nByte;
264 pWInfo->pMemToFree = pBlock;
265 pBlock++;
266 }
267 return (void*)pBlock;
268}
269void *sqlite3WhereRealloc(WhereInfo *pWInfo, void *pOld, u64 nByte){
270 void *pNew = sqlite3WhereMalloc(pWInfo, nByte);
271 if( pNew && pOld ){
272 WhereMemBlock *pOldBlk = (WhereMemBlock*)pOld;
273 pOldBlk--;
274 assert( pOldBlk->sz<nByte );
275 memcpy(pNew, pOld, pOldBlk->sz);
276 }
277 return pNew;
278}
279
drh6a3ea0e2003-05-02 14:32:12 +0000280/*
drh1398ad32005-01-19 23:24:50 +0000281** Create a new mask for cursor iCursor.
drh0fcef5e2005-07-19 17:38:22 +0000282**
283** There is one cursor per table in the FROM clause. The number of
284** tables in the FROM clause is limited by a test early in the
drhb6fb62d2005-09-20 08:47:20 +0000285** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[]
drh0fcef5e2005-07-19 17:38:22 +0000286** array will never overflow.
drh1398ad32005-01-19 23:24:50 +0000287*/
drh111a6a72008-12-21 03:51:16 +0000288static void createMask(WhereMaskSet *pMaskSet, int iCursor){
drhcad651e2007-04-20 12:22:01 +0000289 assert( pMaskSet->n < ArraySize(pMaskSet->ix) );
drh0fcef5e2005-07-19 17:38:22 +0000290 pMaskSet->ix[pMaskSet->n++] = iCursor;
drh1398ad32005-01-19 23:24:50 +0000291}
292
293/*
drh235667a2020-11-08 20:44:30 +0000294** If the right-hand branch of the expression is a TK_COLUMN, then return
295** a pointer to the right-hand branch. Otherwise, return NULL.
296*/
297static Expr *whereRightSubexprIsColumn(Expr *p){
298 p = sqlite3ExprSkipCollateAndLikely(p->pRight);
dan9988db82021-04-15 19:09:19 +0000299 if( ALWAYS(p!=0) && p->op==TK_COLUMN && !ExprHasProperty(p, EP_FixedCol) ){
300 return p;
301 }
drh235667a2020-11-08 20:44:30 +0000302 return 0;
303}
304
305/*
drh1c8148f2013-05-04 20:25:23 +0000306** Advance to the next WhereTerm that matches according to the criteria
307** established when the pScan object was initialized by whereScanInit().
308** Return NULL if there are no more matching WhereTerms.
309*/
danb2cfc142013-07-05 11:10:54 +0000310static WhereTerm *whereScanNext(WhereScan *pScan){
drh1c8148f2013-05-04 20:25:23 +0000311 int iCur; /* The cursor on the LHS of the term */
drha3f108e2015-08-26 21:08:04 +0000312 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */
drh1c8148f2013-05-04 20:25:23 +0000313 Expr *pX; /* An expression being tested */
314 WhereClause *pWC; /* Shorthand for pScan->pWC */
315 WhereTerm *pTerm; /* The term being tested */
drh43b85ef2013-06-10 12:34:45 +0000316 int k = pScan->k; /* Where to start scanning */
drh1c8148f2013-05-04 20:25:23 +0000317
drh392ddeb2016-10-26 17:57:40 +0000318 assert( pScan->iEquiv<=pScan->nEquiv );
319 pWC = pScan->pWC;
320 while(1){
drha3f108e2015-08-26 21:08:04 +0000321 iColumn = pScan->aiColumn[pScan->iEquiv-1];
drh392ddeb2016-10-26 17:57:40 +0000322 iCur = pScan->aiCur[pScan->iEquiv-1];
323 assert( pWC!=0 );
drh220f0d62021-10-15 17:06:16 +0000324 assert( iCur>=0 );
drh392ddeb2016-10-26 17:57:40 +0000325 do{
drh43b85ef2013-06-10 12:34:45 +0000326 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){
drh220f0d62021-10-15 17:06:16 +0000327 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 || pTerm->leftCursor<0 );
drhe1a086e2013-10-28 20:15:56 +0000328 if( pTerm->leftCursor==iCur
drh75fa2662020-09-28 15:49:43 +0000329 && pTerm->u.x.leftColumn==iColumn
drh4b92f982015-09-29 17:20:14 +0000330 && (iColumn!=XN_EXPR
drhf9463df2017-02-11 14:59:58 +0000331 || sqlite3ExprCompareSkip(pTerm->pExpr->pLeft,
332 pScan->pIdxExpr,iCur)==0)
drha3f108e2015-08-26 21:08:04 +0000333 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin))
drhe1a086e2013-10-28 20:15:56 +0000334 ){
drh1c8148f2013-05-04 20:25:23 +0000335 if( (pTerm->eOperator & WO_EQUIV)!=0
drha3f108e2015-08-26 21:08:04 +0000336 && pScan->nEquiv<ArraySize(pScan->aiCur)
drh235667a2020-11-08 20:44:30 +0000337 && (pX = whereRightSubexprIsColumn(pTerm->pExpr))!=0
drh1c8148f2013-05-04 20:25:23 +0000338 ){
339 int j;
drha3f108e2015-08-26 21:08:04 +0000340 for(j=0; j<pScan->nEquiv; j++){
341 if( pScan->aiCur[j]==pX->iTable
342 && pScan->aiColumn[j]==pX->iColumn ){
drh1c8148f2013-05-04 20:25:23 +0000343 break;
344 }
345 }
346 if( j==pScan->nEquiv ){
drha3f108e2015-08-26 21:08:04 +0000347 pScan->aiCur[j] = pX->iTable;
348 pScan->aiColumn[j] = pX->iColumn;
349 pScan->nEquiv++;
drh1c8148f2013-05-04 20:25:23 +0000350 }
351 }
352 if( (pTerm->eOperator & pScan->opMask)!=0 ){
353 /* Verify the affinity and collating sequence match */
354 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){
355 CollSeq *pColl;
drh70d18342013-06-06 19:16:33 +0000356 Parse *pParse = pWC->pWInfo->pParse;
drh1c8148f2013-05-04 20:25:23 +0000357 pX = pTerm->pExpr;
358 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){
359 continue;
360 }
361 assert(pX->pLeft);
drh898c5272019-10-22 00:03:41 +0000362 pColl = sqlite3ExprCompareCollSeq(pParse, pX);
drh70d18342013-06-06 19:16:33 +0000363 if( pColl==0 ) pColl = pParse->db->pDfltColl;
drh1c8148f2013-05-04 20:25:23 +0000364 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){
365 continue;
366 }
367 }
drhe8d0c612015-05-14 01:05:25 +0000368 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0
drhc59ffa82021-10-04 15:08:49 +0000369 && (pX = pTerm->pExpr->pRight, ALWAYS(pX!=0))
370 && pX->op==TK_COLUMN
drha3f108e2015-08-26 21:08:04 +0000371 && pX->iTable==pScan->aiCur[0]
372 && pX->iColumn==pScan->aiColumn[0]
drha184fb82013-05-08 04:22:59 +0000373 ){
drhe8d0c612015-05-14 01:05:25 +0000374 testcase( pTerm->eOperator & WO_IS );
drha184fb82013-05-08 04:22:59 +0000375 continue;
376 }
drh392ddeb2016-10-26 17:57:40 +0000377 pScan->pWC = pWC;
drh43b85ef2013-06-10 12:34:45 +0000378 pScan->k = k+1;
drh6068b6b2021-05-04 16:51:52 +0000379#ifdef WHERETRACE_ENABLED
380 if( sqlite3WhereTrace & 0x20000 ){
381 int ii;
382 sqlite3DebugPrintf("SCAN-TERM %p: nEquiv=%d",
383 pTerm, pScan->nEquiv);
384 for(ii=0; ii<pScan->nEquiv; ii++){
385 sqlite3DebugPrintf(" {%d:%d}",
386 pScan->aiCur[ii], pScan->aiColumn[ii]);
387 }
388 sqlite3DebugPrintf("\n");
389 }
390#endif
drh1c8148f2013-05-04 20:25:23 +0000391 return pTerm;
392 }
393 }
394 }
drh392ddeb2016-10-26 17:57:40 +0000395 pWC = pWC->pOuter;
drh43b85ef2013-06-10 12:34:45 +0000396 k = 0;
drh392ddeb2016-10-26 17:57:40 +0000397 }while( pWC!=0 );
398 if( pScan->iEquiv>=pScan->nEquiv ) break;
399 pWC = pScan->pOrigWC;
drh43b85ef2013-06-10 12:34:45 +0000400 k = 0;
drha3f108e2015-08-26 21:08:04 +0000401 pScan->iEquiv++;
drh1c8148f2013-05-04 20:25:23 +0000402 }
drh1c8148f2013-05-04 20:25:23 +0000403 return 0;
404}
405
406/*
drhe86974c2019-01-28 18:58:54 +0000407** This is whereScanInit() for the case of an index on an expression.
408** It is factored out into a separate tail-recursion subroutine so that
409** the normal whereScanInit() routine, which is a high-runner, does not
410** need to push registers onto the stack as part of its prologue.
411*/
412static SQLITE_NOINLINE WhereTerm *whereScanInitIndexExpr(WhereScan *pScan){
413 pScan->idxaff = sqlite3ExprAffinity(pScan->pIdxExpr);
414 return whereScanNext(pScan);
415}
416
417/*
drh1c8148f2013-05-04 20:25:23 +0000418** Initialize a WHERE clause scanner object. Return a pointer to the
419** first match. Return NULL if there are no matches.
420**
421** The scanner will be searching the WHERE clause pWC. It will look
422** for terms of the form "X <op> <expr>" where X is column iColumn of table
drha3fd75d2016-05-06 18:47:23 +0000423** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx
424** must be one of the indexes of table iCur.
425**
426** The <op> must be one of the operators described by opMask.
drh1c8148f2013-05-04 20:25:23 +0000427**
drh3b48e8c2013-06-12 20:18:16 +0000428** If the search is for X and the WHERE clause contains terms of the
429** form X=Y then this routine might also return terms of the form
430** "Y <op> <expr>". The number of levels of transitivity is limited,
431** but is enough to handle most commonly occurring SQL statements.
432**
drh1c8148f2013-05-04 20:25:23 +0000433** If X is not the INTEGER PRIMARY KEY then X must be compatible with
434** index pIdx.
435*/
danb2cfc142013-07-05 11:10:54 +0000436static WhereTerm *whereScanInit(
drh1c8148f2013-05-04 20:25:23 +0000437 WhereScan *pScan, /* The WhereScan object being initialized */
438 WhereClause *pWC, /* The WHERE clause to be scanned */
439 int iCur, /* Cursor to scan for */
440 int iColumn, /* Column to scan for */
441 u32 opMask, /* Operator(s) to scan for */
442 Index *pIdx /* Must be compatible with this index */
443){
drh1c8148f2013-05-04 20:25:23 +0000444 pScan->pOrigWC = pWC;
445 pScan->pWC = pWC;
drh6860e6f2015-08-27 18:24:02 +0000446 pScan->pIdxExpr = 0;
drh99042982016-10-26 18:41:43 +0000447 pScan->idxaff = 0;
448 pScan->zCollName = 0;
drhe86974c2019-01-28 18:58:54 +0000449 pScan->opMask = opMask;
450 pScan->k = 0;
451 pScan->aiCur[0] = iCur;
452 pScan->nEquiv = 1;
453 pScan->iEquiv = 1;
drhbb523082015-08-27 15:58:51 +0000454 if( pIdx ){
drh99042982016-10-26 18:41:43 +0000455 int j = iColumn;
drhbb523082015-08-27 15:58:51 +0000456 iColumn = pIdx->aiColumn[j];
drh79ab3842021-12-02 02:22:35 +0000457 if( iColumn==pIdx->pTable->iPKey ){
drh99042982016-10-26 18:41:43 +0000458 iColumn = XN_ROWID;
459 }else if( iColumn>=0 ){
460 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity;
461 pScan->zCollName = pIdx->azColl[j];
drh79ab3842021-12-02 02:22:35 +0000462 }else if( iColumn==XN_EXPR ){
463 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr;
464 pScan->zCollName = pIdx->azColl[j];
465 pScan->aiColumn[0] = XN_EXPR;
466 return whereScanInitIndexExpr(pScan);
drh99042982016-10-26 18:41:43 +0000467 }
468 }else if( iColumn==XN_EXPR ){
469 return 0;
drh1c8148f2013-05-04 20:25:23 +0000470 }
drha3f108e2015-08-26 21:08:04 +0000471 pScan->aiColumn[0] = iColumn;
drh1c8148f2013-05-04 20:25:23 +0000472 return whereScanNext(pScan);
473}
474
475/*
drhfe05af82005-07-21 03:14:59 +0000476** Search for a term in the WHERE clause that is of the form "X <op> <expr>"
drha3fd75d2016-05-06 18:47:23 +0000477** where X is a reference to the iColumn of table iCur or of index pIdx
478** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by
479** the op parameter. Return a pointer to the term. Return 0 if not found.
drh58eb1c02013-01-17 00:08:42 +0000480**
drha3fd75d2016-05-06 18:47:23 +0000481** If pIdx!=0 then it must be one of the indexes of table iCur.
482** Search for terms matching the iColumn-th column of pIdx
drhbb523082015-08-27 15:58:51 +0000483** rather than the iColumn-th column of table iCur.
484**
drh58eb1c02013-01-17 00:08:42 +0000485** The term returned might by Y=<expr> if there is another constraint in
486** the WHERE clause that specifies that X=Y. Any such constraints will be
487** identified by the WO_EQUIV bit in the pTerm->eOperator field. The
drha3f108e2015-08-26 21:08:04 +0000488** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11
489** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10
490** other equivalent values. Hence a search for X will return <expr> if X=A1
491** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>.
drh58eb1c02013-01-17 00:08:42 +0000492**
493** If there are multiple terms in the WHERE clause of the form "X <op> <expr>"
494** then try for the one with no dependencies on <expr> - in other words where
495** <expr> is a constant expression of some kind. Only return entries of
496** the form "X <op> Y" where Y is a column in another table if no terms of
drh459f63e2013-03-06 01:55:27 +0000497** the form "X <op> <const-expr>" exist. If no terms with a constant RHS
498** exist, try to return a term that does not use WO_EQUIV.
drhfe05af82005-07-21 03:14:59 +0000499*/
drh6f82e852015-06-06 20:12:09 +0000500WhereTerm *sqlite3WhereFindTerm(
drhfe05af82005-07-21 03:14:59 +0000501 WhereClause *pWC, /* The WHERE clause to be searched */
502 int iCur, /* Cursor number of LHS */
503 int iColumn, /* Column number of LHS */
504 Bitmask notReady, /* RHS must not overlap with this mask */
drhec1724e2008-12-09 01:32:03 +0000505 u32 op, /* Mask of WO_xx values describing operator */
drhfe05af82005-07-21 03:14:59 +0000506 Index *pIdx /* Must be compatible with this index, if not NULL */
507){
drh1c8148f2013-05-04 20:25:23 +0000508 WhereTerm *pResult = 0;
509 WhereTerm *p;
510 WhereScan scan;
drh7a5bcc02013-01-16 17:08:58 +0000511
drh1c8148f2013-05-04 20:25:23 +0000512 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx);
drhe8d0c612015-05-14 01:05:25 +0000513 op &= WO_EQ|WO_IS;
drh1c8148f2013-05-04 20:25:23 +0000514 while( p ){
515 if( (p->prereqRight & notReady)==0 ){
drhe8d0c612015-05-14 01:05:25 +0000516 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){
517 testcase( p->eOperator & WO_IS );
drh1c8148f2013-05-04 20:25:23 +0000518 return p;
drhfe05af82005-07-21 03:14:59 +0000519 }
drh1c8148f2013-05-04 20:25:23 +0000520 if( pResult==0 ) pResult = p;
drhfe05af82005-07-21 03:14:59 +0000521 }
drh1c8148f2013-05-04 20:25:23 +0000522 p = whereScanNext(&scan);
drhfe05af82005-07-21 03:14:59 +0000523 }
drh7a5bcc02013-01-16 17:08:58 +0000524 return pResult;
drhfe05af82005-07-21 03:14:59 +0000525}
526
drh7b4fc6a2007-02-06 13:26:32 +0000527/*
peter.d.reid60ec9142014-09-06 16:39:46 +0000528** This function searches pList for an entry that matches the iCol-th column
drh3b48e8c2013-06-12 20:18:16 +0000529** of index pIdx.
dan6f343962011-07-01 18:26:40 +0000530**
531** If such an expression is found, its index in pList->a[] is returned. If
532** no expression is found, -1 is returned.
533*/
534static int findIndexCol(
535 Parse *pParse, /* Parse context */
536 ExprList *pList, /* Expression list to search */
537 int iBase, /* Cursor for table associated with pIdx */
538 Index *pIdx, /* Index to match column of */
539 int iCol /* Column of index to match */
540){
541 int i;
542 const char *zColl = pIdx->azColl[iCol];
543
544 for(i=0; i<pList->nExpr; i++){
drh0d950af2019-08-22 16:38:42 +0000545 Expr *p = sqlite3ExprSkipCollateAndLikely(pList->a[i].pExpr);
drh235667a2020-11-08 20:44:30 +0000546 if( ALWAYS(p!=0)
dan4fcb30b2021-03-09 16:06:25 +0000547 && (p->op==TK_COLUMN || p->op==TK_AGG_COLUMN)
drhf1d3e322011-07-09 13:00:41 +0000548 && p->iColumn==pIdx->aiColumn[iCol]
549 && p->iTable==iBase
550 ){
drh70efa842017-09-28 01:58:23 +0000551 CollSeq *pColl = sqlite3ExprNNCollSeq(pParse, pList->a[i].pExpr);
552 if( 0==sqlite3StrICmp(pColl->zName, zColl) ){
dan6f343962011-07-01 18:26:40 +0000553 return i;
554 }
555 }
556 }
557
558 return -1;
559}
560
561/*
drhbb523082015-08-27 15:58:51 +0000562** Return TRUE if the iCol-th column of index pIdx is NOT NULL
563*/
564static int indexColumnNotNull(Index *pIdx, int iCol){
565 int j;
566 assert( pIdx!=0 );
567 assert( iCol>=0 && iCol<pIdx->nColumn );
568 j = pIdx->aiColumn[iCol];
569 if( j>=0 ){
570 return pIdx->pTable->aCol[j].notNull;
571 }else if( j==(-1) ){
572 return 1;
573 }else{
574 assert( j==(-2) );
drh84926532015-08-31 19:38:42 +0000575 return 0; /* Assume an indexed expression can always yield a NULL */
drh7d3d9da2015-09-01 00:42:52 +0000576
drhbb523082015-08-27 15:58:51 +0000577 }
578}
579
580/*
dan6f343962011-07-01 18:26:40 +0000581** Return true if the DISTINCT expression-list passed as the third argument
drh4f402f22013-06-11 18:59:38 +0000582** is redundant.
583**
drhb121dd12015-06-06 18:30:17 +0000584** A DISTINCT list is redundant if any subset of the columns in the
585** DISTINCT list are collectively unique and individually non-null.
dan6f343962011-07-01 18:26:40 +0000586*/
587static int isDistinctRedundant(
drh4f402f22013-06-11 18:59:38 +0000588 Parse *pParse, /* Parsing context */
589 SrcList *pTabList, /* The FROM clause */
590 WhereClause *pWC, /* The WHERE clause */
591 ExprList *pDistinct /* The result set that needs to be DISTINCT */
dan6f343962011-07-01 18:26:40 +0000592){
593 Table *pTab;
594 Index *pIdx;
595 int i;
596 int iBase;
597
598 /* If there is more than one table or sub-select in the FROM clause of
599 ** this query, then it will not be possible to show that the DISTINCT
600 ** clause is redundant. */
601 if( pTabList->nSrc!=1 ) return 0;
602 iBase = pTabList->a[0].iCursor;
603 pTab = pTabList->a[0].pTab;
604
dan94e08d92011-07-02 06:44:05 +0000605 /* If any of the expressions is an IPK column on table iBase, then return
606 ** true. Note: The (p->iTable==iBase) part of this test may be false if the
607 ** current SELECT is a correlated sub-query.
608 */
dan6f343962011-07-01 18:26:40 +0000609 for(i=0; i<pDistinct->nExpr; i++){
drh0d950af2019-08-22 16:38:42 +0000610 Expr *p = sqlite3ExprSkipCollateAndLikely(pDistinct->a[i].pExpr);
drh235667a2020-11-08 20:44:30 +0000611 if( NEVER(p==0) ) continue;
dan4fcb30b2021-03-09 16:06:25 +0000612 if( p->op!=TK_COLUMN && p->op!=TK_AGG_COLUMN ) continue;
613 if( p->iTable==iBase && p->iColumn<0 ) return 1;
dan6f343962011-07-01 18:26:40 +0000614 }
615
616 /* Loop through all indices on the table, checking each to see if it makes
617 ** the DISTINCT qualifier redundant. It does so if:
618 **
619 ** 1. The index is itself UNIQUE, and
620 **
621 ** 2. All of the columns in the index are either part of the pDistinct
622 ** list, or else the WHERE clause contains a term of the form "col=X",
623 ** where X is a constant value. The collation sequences of the
624 ** comparison and select-list expressions must match those of the index.
dan6a36f432012-04-20 16:59:24 +0000625 **
626 ** 3. All of those index columns for which the WHERE clause does not
627 ** contain a "col=X" term are subject to a NOT NULL constraint.
dan6f343962011-07-01 18:26:40 +0000628 */
629 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){
drh5f1d1d92014-07-31 22:59:04 +0000630 if( !IsUniqueIndex(pIdx) ) continue;
drh204b6342021-04-06 23:29:41 +0000631 if( pIdx->pPartIdxWhere ) continue;
drhbbbdc832013-10-22 18:01:40 +0000632 for(i=0; i<pIdx->nKeyCol; i++){
drhbb523082015-08-27 15:58:51 +0000633 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){
634 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break;
635 if( indexColumnNotNull(pIdx, i)==0 ) break;
dan6f343962011-07-01 18:26:40 +0000636 }
637 }
drhbbbdc832013-10-22 18:01:40 +0000638 if( i==pIdx->nKeyCol ){
dan6f343962011-07-01 18:26:40 +0000639 /* This index implies that the DISTINCT qualifier is redundant. */
640 return 1;
641 }
642 }
643
644 return 0;
645}
drh0fcef5e2005-07-19 17:38:22 +0000646
drh8636e9c2013-06-11 01:50:08 +0000647
drh75897232000-05-29 14:26:00 +0000648/*
drh3b48e8c2013-06-12 20:18:16 +0000649** Estimate the logarithm of the input value to base 2.
drh28c4cf42005-07-27 20:41:43 +0000650*/
drhbf539c42013-10-05 18:16:02 +0000651static LogEst estLog(LogEst N){
drh696964d2014-06-12 15:46:46 +0000652 return N<=10 ? 0 : sqlite3LogEst(N) - 33;
drh28c4cf42005-07-27 20:41:43 +0000653}
654
drh6d209d82006-06-27 01:54:26 +0000655/*
drh7b3aa082015-05-29 13:55:33 +0000656** Convert OP_Column opcodes to OP_Copy in previously generated code.
657**
658** This routine runs over generated VDBE code and translates OP_Column
danfb785b22015-10-24 20:31:22 +0000659** opcodes into OP_Copy when the table is being accessed via co-routine
660** instead of via table lookup.
661**
drh00a61532019-06-28 07:08:13 +0000662** If the iAutoidxCur is not zero, then any OP_Rowid instructions on
663** cursor iTabCur are transformed into OP_Sequence opcode for the
664** iAutoidxCur cursor, in order to generate unique rowids for the
665** automatic index being generated.
drh7b3aa082015-05-29 13:55:33 +0000666*/
667static void translateColumnToCopy(
drh202230e2017-03-11 13:02:59 +0000668 Parse *pParse, /* Parsing context */
drh7b3aa082015-05-29 13:55:33 +0000669 int iStart, /* Translate from this opcode to the end */
670 int iTabCur, /* OP_Column/OP_Rowid references to this table */
danfb785b22015-10-24 20:31:22 +0000671 int iRegister, /* The first column is in this register */
drh00a61532019-06-28 07:08:13 +0000672 int iAutoidxCur /* If non-zero, cursor of autoindex being generated */
drh7b3aa082015-05-29 13:55:33 +0000673){
drh202230e2017-03-11 13:02:59 +0000674 Vdbe *v = pParse->pVdbe;
drh7b3aa082015-05-29 13:55:33 +0000675 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart);
676 int iEnd = sqlite3VdbeCurrentAddr(v);
drh202230e2017-03-11 13:02:59 +0000677 if( pParse->db->mallocFailed ) return;
drh7b3aa082015-05-29 13:55:33 +0000678 for(; iStart<iEnd; iStart++, pOp++){
679 if( pOp->p1!=iTabCur ) continue;
680 if( pOp->opcode==OP_Column ){
681 pOp->opcode = OP_Copy;
682 pOp->p1 = pOp->p2 + iRegister;
683 pOp->p2 = pOp->p3;
684 pOp->p3 = 0;
685 }else if( pOp->opcode==OP_Rowid ){
drh6e5020e2021-04-07 15:45:01 +0000686 pOp->opcode = OP_Sequence;
687 pOp->p1 = iAutoidxCur;
688#ifdef SQLITE_ALLOW_ROWID_IN_VIEW
689 if( iAutoidxCur==0 ){
danfb785b22015-10-24 20:31:22 +0000690 pOp->opcode = OP_Null;
danfb785b22015-10-24 20:31:22 +0000691 pOp->p3 = 0;
692 }
drh6e5020e2021-04-07 15:45:01 +0000693#endif
drh7b3aa082015-05-29 13:55:33 +0000694 }
695 }
696}
697
698/*
drh6d209d82006-06-27 01:54:26 +0000699** Two routines for printing the content of an sqlite3_index_info
700** structure. Used for testing and debugging only. If neither
701** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines
702** are no-ops.
703*/
drhd15cb172013-05-21 19:23:10 +0000704#if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED)
drhcfcf4de2019-12-28 13:01:52 +0000705static void whereTraceIndexInfoInputs(sqlite3_index_info *p){
drh6d209d82006-06-27 01:54:26 +0000706 int i;
mlcreech3a00f902008-03-04 17:45:01 +0000707 if( !sqlite3WhereTrace ) return;
drh6d209d82006-06-27 01:54:26 +0000708 for(i=0; i<p->nConstraint; i++){
drhb6592f62021-12-17 23:56:43 +0000709 sqlite3DebugPrintf(
710 " constraint[%d]: col=%d termid=%d op=%d usabled=%d collseq=%s\n",
drh6d209d82006-06-27 01:54:26 +0000711 i,
712 p->aConstraint[i].iColumn,
713 p->aConstraint[i].iTermOffset,
714 p->aConstraint[i].op,
drhb6592f62021-12-17 23:56:43 +0000715 p->aConstraint[i].usable,
716 sqlite3_vtab_collation(p,i));
drh6d209d82006-06-27 01:54:26 +0000717 }
718 for(i=0; i<p->nOrderBy; i++){
719 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n",
720 i,
721 p->aOrderBy[i].iColumn,
722 p->aOrderBy[i].desc);
723 }
724}
drhcfcf4de2019-12-28 13:01:52 +0000725static void whereTraceIndexInfoOutputs(sqlite3_index_info *p){
drh6d209d82006-06-27 01:54:26 +0000726 int i;
mlcreech3a00f902008-03-04 17:45:01 +0000727 if( !sqlite3WhereTrace ) return;
drh6d209d82006-06-27 01:54:26 +0000728 for(i=0; i<p->nConstraint; i++){
729 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n",
730 i,
731 p->aConstraintUsage[i].argvIndex,
732 p->aConstraintUsage[i].omit);
733 }
734 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum);
735 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr);
736 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed);
737 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost);
dana9f58152013-11-11 19:01:33 +0000738 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows);
drh6d209d82006-06-27 01:54:26 +0000739}
740#else
drhcfcf4de2019-12-28 13:01:52 +0000741#define whereTraceIndexInfoInputs(A)
742#define whereTraceIndexInfoOutputs(A)
drh6d209d82006-06-27 01:54:26 +0000743#endif
744
drhc6339082010-04-07 16:54:58 +0000745#ifndef SQLITE_OMIT_AUTOMATIC_INDEX
drh8b307fb2010-04-06 15:57:05 +0000746/*
drh4139c992010-04-07 14:59:45 +0000747** Return TRUE if the WHERE clause term pTerm is of a form where it
748** could be used with an index to access pSrc, assuming an appropriate
749** index existed.
750*/
751static int termCanDriveIndex(
drhfecbf0a2021-12-04 21:11:18 +0000752 const WhereTerm *pTerm, /* WHERE clause term to check */
753 const SrcItem *pSrc, /* Table we are trying to access */
754 const Bitmask notReady /* Tables in outer loops of the join */
drh4139c992010-04-07 14:59:45 +0000755){
756 char aff;
757 if( pTerm->leftCursor!=pSrc->iCursor ) return 0;
drhe8d0c612015-05-14 01:05:25 +0000758 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0;
drha76ac882022-04-08 19:20:12 +0000759 if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ))!=0
danbbccd522017-07-18 17:13:41 +0000760 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin)
761 && (pTerm->eOperator & WO_IS)
762 ){
763 /* Cannot use an IS term from the WHERE clause as an index driver for
drha76ac882022-04-08 19:20:12 +0000764 ** the RHS of a LEFT JOIN or for the LHS of a RIGHT JOIN. Such a term
765 ** can only be used if it is from the ON clause. */
danbbccd522017-07-18 17:13:41 +0000766 return 0;
767 }
drh4139c992010-04-07 14:59:45 +0000768 if( (pTerm->prereqRight & notReady)!=0 ) return 0;
drh220f0d62021-10-15 17:06:16 +0000769 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 );
drh75fa2662020-09-28 15:49:43 +0000770 if( pTerm->u.x.leftColumn<0 ) return 0;
771 aff = pSrc->pTab->aCol[pTerm->u.x.leftColumn].affinity;
drh4139c992010-04-07 14:59:45 +0000772 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0;
drhe0cc3c22015-05-13 17:54:08 +0000773 testcase( pTerm->pExpr->op==TK_IS );
drh4139c992010-04-07 14:59:45 +0000774 return 1;
775}
drhc6339082010-04-07 16:54:58 +0000776#endif
drh4139c992010-04-07 14:59:45 +0000777
drhc6339082010-04-07 16:54:58 +0000778
779#ifndef SQLITE_OMIT_AUTOMATIC_INDEX
drh8b307fb2010-04-06 15:57:05 +0000780/*
drhc6339082010-04-07 16:54:58 +0000781** Generate code to construct the Index object for an automatic index
782** and to set up the WhereLevel object pLevel so that the code generator
783** makes use of the automatic index.
drh8b307fb2010-04-06 15:57:05 +0000784*/
drhfa35f5c2021-12-04 13:43:57 +0000785static SQLITE_NOINLINE void constructAutomaticIndex(
drh8b307fb2010-04-06 15:57:05 +0000786 Parse *pParse, /* The parsing context */
drhfecbf0a2021-12-04 21:11:18 +0000787 const WhereClause *pWC, /* The WHERE clause */
788 const SrcItem *pSrc, /* The FROM clause term to get the next index */
789 const Bitmask notReady, /* Mask of cursors that are not available */
drh8b307fb2010-04-06 15:57:05 +0000790 WhereLevel *pLevel /* Write new index here */
791){
drhbbbdc832013-10-22 18:01:40 +0000792 int nKeyCol; /* Number of columns in the constructed index */
drh8b307fb2010-04-06 15:57:05 +0000793 WhereTerm *pTerm; /* A single term of the WHERE clause */
794 WhereTerm *pWCEnd; /* End of pWC->a[] */
drh8b307fb2010-04-06 15:57:05 +0000795 Index *pIdx; /* Object describing the transient index */
796 Vdbe *v; /* Prepared statement under construction */
drh8b307fb2010-04-06 15:57:05 +0000797 int addrInit; /* Address of the initialization bypass jump */
798 Table *pTable; /* The table being indexed */
drh8b307fb2010-04-06 15:57:05 +0000799 int addrTop; /* Top of the index fill loop */
800 int regRecord; /* Register holding an index record */
801 int n; /* Column counter */
drh4139c992010-04-07 14:59:45 +0000802 int i; /* Loop counter */
803 int mxBitCol; /* Maximum column in pSrc->colUsed */
drh424aab82010-04-06 18:28:20 +0000804 CollSeq *pColl; /* Collating sequence to on a column */
drh7ba39a92013-05-30 17:43:19 +0000805 WhereLoop *pLoop; /* The Loop object */
drh77e57df2013-10-22 14:28:02 +0000806 char *zNotUsed; /* Extra space on the end of pIdx */
drh4139c992010-04-07 14:59:45 +0000807 Bitmask idxCols; /* Bitmap of columns used for indexing */
808 Bitmask extraCols; /* Bitmap of additional columns */
drh8d56e202013-06-28 23:55:45 +0000809 u8 sentWarning = 0; /* True if a warnning has been issued */
drh059b2d52014-10-24 19:28:09 +0000810 Expr *pPartial = 0; /* Partial Index Expression */
811 int iContinue = 0; /* Jump here to skip excluded rows */
drh76012942021-02-21 21:04:54 +0000812 SrcItem *pTabItem; /* FROM clause term being indexed */
drh4dd83a22015-10-26 14:54:32 +0000813 int addrCounter = 0; /* Address where integer counter is initialized */
danfb785b22015-10-24 20:31:22 +0000814 int regBase; /* Array of registers where record is assembled */
drh8b307fb2010-04-06 15:57:05 +0000815
816 /* Generate code to skip over the creation and initialization of the
817 ** transient index on 2nd and subsequent iterations of the loop. */
818 v = pParse->pVdbe;
819 assert( v!=0 );
drh511f9e82016-09-22 18:53:13 +0000820 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v);
drh8b307fb2010-04-06 15:57:05 +0000821
drh4139c992010-04-07 14:59:45 +0000822 /* Count the number of columns that will be added to the index
823 ** and used to match WHERE clause constraints */
drhbbbdc832013-10-22 18:01:40 +0000824 nKeyCol = 0;
drh424aab82010-04-06 18:28:20 +0000825 pTable = pSrc->pTab;
drh8b307fb2010-04-06 15:57:05 +0000826 pWCEnd = &pWC->a[pWC->nTerm];
drh7ba39a92013-05-30 17:43:19 +0000827 pLoop = pLevel->pWLoop;
drh4139c992010-04-07 14:59:45 +0000828 idxCols = 0;
drh81186b42013-06-18 01:52:41 +0000829 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
drh13cc90c2015-02-25 00:24:41 +0000830 Expr *pExpr = pTerm->pExpr;
drhc1085ea2021-11-30 14:07:58 +0000831 /* Make the automatic index a partial index if there are terms in the
832 ** WHERE clause (or the ON clause of a LEFT join) that constrain which
833 ** rows of the target table (pSrc) that can be used. */
834 if( (pTerm->wtFlags & TERM_VIRTUAL)==0
drha76ac882022-04-08 19:20:12 +0000835 && ((pSrc->fg.jointype&(JT_LEFT|JT_LTORJ))==0
836 || ExprHasProperty(pExpr,EP_FromJoin))
drhc1085ea2021-11-30 14:07:58 +0000837 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor)
838 ){
drhd5c851c2019-04-19 13:38:34 +0000839 pPartial = sqlite3ExprAnd(pParse, pPartial,
drh13cc90c2015-02-25 00:24:41 +0000840 sqlite3ExprDup(pParse->db, pExpr, 0));
drh059b2d52014-10-24 19:28:09 +0000841 }
drh4139c992010-04-07 14:59:45 +0000842 if( termCanDriveIndex(pTerm, pSrc, notReady) ){
drh220f0d62021-10-15 17:06:16 +0000843 int iCol;
844 Bitmask cMask;
845 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 );
846 iCol = pTerm->u.x.leftColumn;
847 cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol);
drh52ff8ea2010-04-08 14:15:56 +0000848 testcase( iCol==BMS );
849 testcase( iCol==BMS-1 );
drh8d56e202013-06-28 23:55:45 +0000850 if( !sentWarning ){
851 sqlite3_log(SQLITE_WARNING_AUTOINDEX,
852 "automatic index on %s(%s)", pTable->zName,
drhcf9d36d2021-08-02 18:03:43 +0000853 pTable->aCol[iCol].zCnName);
drh8d56e202013-06-28 23:55:45 +0000854 sentWarning = 1;
855 }
drh0013e722010-04-08 00:40:15 +0000856 if( (idxCols & cMask)==0 ){
drh059b2d52014-10-24 19:28:09 +0000857 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){
858 goto end_auto_index_create;
859 }
drhbbbdc832013-10-22 18:01:40 +0000860 pLoop->aLTerm[nKeyCol++] = pTerm;
drh0013e722010-04-08 00:40:15 +0000861 idxCols |= cMask;
862 }
drh8b307fb2010-04-06 15:57:05 +0000863 }
864 }
drhd6a33de2021-04-07 12:36:58 +0000865 assert( nKeyCol>0 || pParse->db->mallocFailed );
drhbbbdc832013-10-22 18:01:40 +0000866 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol;
drh53b52f72013-05-31 11:57:39 +0000867 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED
drh986b3872013-06-28 21:12:20 +0000868 | WHERE_AUTO_INDEX;
drh4139c992010-04-07 14:59:45 +0000869
870 /* Count the number of additional columns needed to create a
871 ** covering index. A "covering index" is an index that contains all
872 ** columns that are needed by the query. With a covering index, the
873 ** original table never needs to be accessed. Automatic indices must
874 ** be a covering index because the index will not be updated if the
875 ** original table changes and the index and table cannot both be used
876 ** if they go out of sync.
877 */
drh7699d1c2013-06-04 12:42:29 +0000878 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1));
drhc3ef4fa2014-10-28 15:58:50 +0000879 mxBitCol = MIN(BMS-1,pTable->nCol);
drh52ff8ea2010-04-08 14:15:56 +0000880 testcase( pTable->nCol==BMS-1 );
881 testcase( pTable->nCol==BMS-2 );
drh4139c992010-04-07 14:59:45 +0000882 for(i=0; i<mxBitCol; i++){
drhbbbdc832013-10-22 18:01:40 +0000883 if( extraCols & MASKBIT(i) ) nKeyCol++;
drh4139c992010-04-07 14:59:45 +0000884 }
drh7699d1c2013-06-04 12:42:29 +0000885 if( pSrc->colUsed & MASKBIT(BMS-1) ){
drhbbbdc832013-10-22 18:01:40 +0000886 nKeyCol += pTable->nCol - BMS + 1;
drh4139c992010-04-07 14:59:45 +0000887 }
drh8b307fb2010-04-06 15:57:05 +0000888
889 /* Construct the Index object to describe this index */
drhbbbdc832013-10-22 18:01:40 +0000890 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed);
drh059b2d52014-10-24 19:28:09 +0000891 if( pIdx==0 ) goto end_auto_index_create;
drh7ba39a92013-05-30 17:43:19 +0000892 pLoop->u.btree.pIndex = pIdx;
drh8b307fb2010-04-06 15:57:05 +0000893 pIdx->zName = "auto-index";
drh424aab82010-04-06 18:28:20 +0000894 pIdx->pTable = pTable;
drh8b307fb2010-04-06 15:57:05 +0000895 n = 0;
drh0013e722010-04-08 00:40:15 +0000896 idxCols = 0;
drh8b307fb2010-04-06 15:57:05 +0000897 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
drh4139c992010-04-07 14:59:45 +0000898 if( termCanDriveIndex(pTerm, pSrc, notReady) ){
drh220f0d62021-10-15 17:06:16 +0000899 int iCol;
900 Bitmask cMask;
901 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 );
902 iCol = pTerm->u.x.leftColumn;
903 cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol);
drh7963b0e2013-06-17 21:37:40 +0000904 testcase( iCol==BMS-1 );
905 testcase( iCol==BMS );
drh0013e722010-04-08 00:40:15 +0000906 if( (idxCols & cMask)==0 ){
907 Expr *pX = pTerm->pExpr;
908 idxCols |= cMask;
drh75fa2662020-09-28 15:49:43 +0000909 pIdx->aiColumn[n] = pTerm->u.x.leftColumn;
drh898c5272019-10-22 00:03:41 +0000910 pColl = sqlite3ExprCompareCollSeq(pParse, pX);
drh34da2a42019-12-24 13:41:33 +0000911 assert( pColl!=0 || pParse->nErr>0 ); /* TH3 collate01.800 */
912 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY;
drh0013e722010-04-08 00:40:15 +0000913 n++;
914 }
drh8b307fb2010-04-06 15:57:05 +0000915 }
916 }
drh7ba39a92013-05-30 17:43:19 +0000917 assert( (u32)n==pLoop->u.btree.nEq );
drh4139c992010-04-07 14:59:45 +0000918
drhc6339082010-04-07 16:54:58 +0000919 /* Add additional columns needed to make the automatic index into
920 ** a covering index */
drh4139c992010-04-07 14:59:45 +0000921 for(i=0; i<mxBitCol; i++){
drh7699d1c2013-06-04 12:42:29 +0000922 if( extraCols & MASKBIT(i) ){
drh4139c992010-04-07 14:59:45 +0000923 pIdx->aiColumn[n] = i;
drhf19aa5f2015-12-30 16:51:20 +0000924 pIdx->azColl[n] = sqlite3StrBINARY;
drh4139c992010-04-07 14:59:45 +0000925 n++;
926 }
927 }
drh7699d1c2013-06-04 12:42:29 +0000928 if( pSrc->colUsed & MASKBIT(BMS-1) ){
drh4139c992010-04-07 14:59:45 +0000929 for(i=BMS-1; i<pTable->nCol; i++){
930 pIdx->aiColumn[n] = i;
drhf19aa5f2015-12-30 16:51:20 +0000931 pIdx->azColl[n] = sqlite3StrBINARY;
drh4139c992010-04-07 14:59:45 +0000932 n++;
933 }
934 }
drhbbbdc832013-10-22 18:01:40 +0000935 assert( n==nKeyCol );
drh4b92f982015-09-29 17:20:14 +0000936 pIdx->aiColumn[n] = XN_ROWID;
drhf19aa5f2015-12-30 16:51:20 +0000937 pIdx->azColl[n] = sqlite3StrBINARY;
drh8b307fb2010-04-06 15:57:05 +0000938
drhc6339082010-04-07 16:54:58 +0000939 /* Create the automatic index */
drh8b307fb2010-04-06 15:57:05 +0000940 assert( pLevel->iIdxCur>=0 );
drha1f41242013-05-31 20:00:58 +0000941 pLevel->iIdxCur = pParse->nTab++;
drh2ec2fb22013-11-06 19:59:23 +0000942 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1);
943 sqlite3VdbeSetP4KeyInfo(pParse, pIdx);
drha21a64d2010-04-06 22:33:55 +0000944 VdbeComment((v, "for %s", pTable->zName));
drh2db144c2021-12-01 16:31:02 +0000945 if( OptimizationEnabled(pParse->db, SQLITE_BloomFilter) ){
946 pLevel->regFilter = ++pParse->nMem;
drh50fb7e02021-12-06 20:16:53 +0000947 sqlite3VdbeAddOp2(v, OP_Blob, 10000, pLevel->regFilter);
drh2db144c2021-12-01 16:31:02 +0000948 }
drh8b307fb2010-04-06 15:57:05 +0000949
drhc6339082010-04-07 16:54:58 +0000950 /* Fill the automatic index with content */
drh7b3aa082015-05-29 13:55:33 +0000951 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom];
drh8a48b9c2015-08-19 15:20:00 +0000952 if( pTabItem->fg.viaCoroutine ){
drh7b3aa082015-05-29 13:55:33 +0000953 int regYield = pTabItem->regReturn;
danfb785b22015-10-24 20:31:22 +0000954 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0);
drh7b3aa082015-05-29 13:55:33 +0000955 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub);
956 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield);
957 VdbeCoverage(v);
drhfef37762018-07-10 19:48:35 +0000958 VdbeComment((v, "next row of %s", pTabItem->pTab->zName));
drh7b3aa082015-05-29 13:55:33 +0000959 }else{
960 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v);
961 }
drh059b2d52014-10-24 19:28:09 +0000962 if( pPartial ){
drhec4ccdb2018-12-29 02:26:59 +0000963 iContinue = sqlite3VdbeMakeLabel(pParse);
drh059b2d52014-10-24 19:28:09 +0000964 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL);
drh051575c2014-10-25 12:28:25 +0000965 pLoop->wsFlags |= WHERE_PARTIALIDX;
drh059b2d52014-10-24 19:28:09 +0000966 }
drh8b307fb2010-04-06 15:57:05 +0000967 regRecord = sqlite3GetTempReg(pParse);
danfb785b22015-10-24 20:31:22 +0000968 regBase = sqlite3GenerateIndexKey(
969 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0
970 );
drh2db144c2021-12-01 16:31:02 +0000971 if( pLevel->regFilter ){
972 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0,
973 regBase, pLoop->u.btree.nEq);
974 }
drh8b307fb2010-04-06 15:57:05 +0000975 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord);
976 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT);
drh059b2d52014-10-24 19:28:09 +0000977 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue);
drh8a48b9c2015-08-19 15:20:00 +0000978 if( pTabItem->fg.viaCoroutine ){
danfb785b22015-10-24 20:31:22 +0000979 sqlite3VdbeChangeP2(v, addrCounter, regBase+n);
drh202230e2017-03-11 13:02:59 +0000980 testcase( pParse->db->mallocFailed );
drh00a61532019-06-28 07:08:13 +0000981 assert( pLevel->iIdxCur>0 );
drh202230e2017-03-11 13:02:59 +0000982 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur,
drh00a61532019-06-28 07:08:13 +0000983 pTabItem->regResult, pLevel->iIdxCur);
drh076e85f2015-09-03 13:46:12 +0000984 sqlite3VdbeGoto(v, addrTop);
drh2c041312018-12-24 02:34:49 +0000985 pTabItem->fg.viaCoroutine = 0;
drh7b3aa082015-05-29 13:55:33 +0000986 }else{
987 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v);
drhd9670ab2019-12-28 01:52:46 +0000988 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX);
drh7b3aa082015-05-29 13:55:33 +0000989 }
drh8b307fb2010-04-06 15:57:05 +0000990 sqlite3VdbeJumpHere(v, addrTop);
991 sqlite3ReleaseTempReg(pParse, regRecord);
992
993 /* Jump here when skipping the initialization */
994 sqlite3VdbeJumpHere(v, addrInit);
drh059b2d52014-10-24 19:28:09 +0000995
996end_auto_index_create:
997 sqlite3ExprDelete(pParse->db, pPartial);
drh8b307fb2010-04-06 15:57:05 +0000998}
drhc6339082010-04-07 16:54:58 +0000999#endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
drh8b307fb2010-04-06 15:57:05 +00001000
drhfa35f5c2021-12-04 13:43:57 +00001001/*
drh6ae49e62021-12-05 20:19:47 +00001002** Generate bytecode that will initialize a Bloom filter that is appropriate
1003** for pLevel.
1004**
1005** If there are inner loops within pLevel that have the WHERE_BLOOMFILTER
1006** flag set, initialize a Bloomfilter for them as well. Except don't do
1007** this recursive initialization if the SQLITE_BloomPulldown optimization has
1008** been turned off.
1009**
1010** When the Bloom filter is initialized, the WHERE_BLOOMFILTER flag is cleared
1011** from the loop, but the regFilter value is set to a register that implements
1012** the Bloom filter. When regFilter is positive, the
1013** sqlite3WhereCodeOneLoopStart() will generate code to test the Bloom filter
1014** and skip the subsequence B-Tree seek if the Bloom filter indicates that
1015** no matching rows exist.
1016**
1017** This routine may only be called if it has previously been determined that
1018** the loop would benefit from a Bloom filter, and the WHERE_BLOOMFILTER bit
1019** is set.
drhfa35f5c2021-12-04 13:43:57 +00001020*/
drh27a9e1f2021-12-10 17:36:16 +00001021static SQLITE_NOINLINE void sqlite3ConstructBloomFilter(
drh6ae49e62021-12-05 20:19:47 +00001022 WhereInfo *pWInfo, /* The WHERE clause */
1023 int iLevel, /* Index in pWInfo->a[] that is pLevel */
drh761d64b2021-12-07 22:37:50 +00001024 WhereLevel *pLevel, /* Make a Bloom filter for this FROM term */
1025 Bitmask notReady /* Loops that are not ready */
drhfa35f5c2021-12-04 13:43:57 +00001026){
drh6ae49e62021-12-05 20:19:47 +00001027 int addrOnce; /* Address of opening OP_Once */
1028 int addrTop; /* Address of OP_Rewind */
1029 int addrCont; /* Jump here to skip a row */
1030 const WhereTerm *pTerm; /* For looping over WHERE clause terms */
1031 const WhereTerm *pWCEnd; /* Last WHERE clause term */
1032 Parse *pParse = pWInfo->pParse; /* Parsing context */
1033 Vdbe *v = pParse->pVdbe; /* VDBE under construction */
1034 WhereLoop *pLoop = pLevel->pWLoop; /* The loop being coded */
1035 int iCur; /* Cursor for table getting the filter */
drhfa35f5c2021-12-04 13:43:57 +00001036
1037 assert( pLoop!=0 );
1038 assert( v!=0 );
drh6ae49e62021-12-05 20:19:47 +00001039 assert( pLoop->wsFlags & WHERE_BLOOMFILTER );
1040
1041 addrOnce = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v);
1042 do{
drh50fb7e02021-12-06 20:16:53 +00001043 const SrcItem *pItem;
1044 const Table *pTab;
drh7e910f62021-12-09 01:28:15 +00001045 u64 sz;
drh6ae49e62021-12-05 20:19:47 +00001046 sqlite3WhereExplainBloomFilter(pParse, pWInfo, pLevel);
1047 addrCont = sqlite3VdbeMakeLabel(pParse);
1048 iCur = pLevel->iTabCur;
1049 pLevel->regFilter = ++pParse->nMem;
drh50fb7e02021-12-06 20:16:53 +00001050
1051 /* The Bloom filter is a Blob held in a register. Initialize it
1052 ** to zero-filled blob of at least 80K bits, but maybe more if the
1053 ** estimated size of the table is larger. We could actually
1054 ** measure the size of the table at run-time using OP_Count with
1055 ** P3==1 and use that value to initialize the blob. But that makes
1056 ** testing complicated. By basing the blob size on the value in the
1057 ** sqlite_stat1 table, testing is much easier.
1058 */
1059 pItem = &pWInfo->pTabList->a[pLevel->iFrom];
1060 assert( pItem!=0 );
1061 pTab = pItem->pTab;
1062 assert( pTab!=0 );
drh7e910f62021-12-09 01:28:15 +00001063 sz = sqlite3LogEstToInt(pTab->nRowLogEst);
1064 if( sz<10000 ){
drh50fb7e02021-12-06 20:16:53 +00001065 sz = 10000;
drh7e910f62021-12-09 01:28:15 +00001066 }else if( sz>10000000 ){
1067 sz = 10000000;
drh50fb7e02021-12-06 20:16:53 +00001068 }
drh7e910f62021-12-09 01:28:15 +00001069 sqlite3VdbeAddOp2(v, OP_Blob, (int)sz, pLevel->regFilter);
drh50fb7e02021-12-06 20:16:53 +00001070
drh6ae49e62021-12-05 20:19:47 +00001071 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, iCur); VdbeCoverage(v);
1072 pWCEnd = &pWInfo->sWC.a[pWInfo->sWC.nTerm];
1073 for(pTerm=pWInfo->sWC.a; pTerm<pWCEnd; pTerm++){
1074 Expr *pExpr = pTerm->pExpr;
1075 if( (pTerm->wtFlags & TERM_VIRTUAL)==0
1076 && sqlite3ExprIsTableConstant(pExpr, iCur)
1077 ){
1078 sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL);
1079 }
drhfa35f5c2021-12-04 13:43:57 +00001080 }
drh6ae49e62021-12-05 20:19:47 +00001081 if( pLoop->wsFlags & WHERE_IPK ){
1082 int r1 = sqlite3GetTempReg(pParse);
1083 sqlite3VdbeAddOp2(v, OP_Rowid, iCur, r1);
1084 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0, r1, 1);
1085 sqlite3ReleaseTempReg(pParse, r1);
1086 }else{
1087 Index *pIdx = pLoop->u.btree.pIndex;
1088 int n = pLoop->u.btree.nEq;
1089 int r1 = sqlite3GetTempRange(pParse, n);
1090 int jj;
1091 for(jj=0; jj<n; jj++){
1092 int iCol = pIdx->aiColumn[jj];
drh50fb7e02021-12-06 20:16:53 +00001093 assert( pIdx->pTable==pItem->pTab );
drh6ae49e62021-12-05 20:19:47 +00001094 sqlite3ExprCodeGetColumnOfTable(v, pIdx->pTable, iCur, iCol,r1+jj);
1095 }
1096 sqlite3VdbeAddOp4Int(v, OP_FilterAdd, pLevel->regFilter, 0, r1, n);
1097 sqlite3ReleaseTempRange(pParse, r1, n);
drhfa35f5c2021-12-04 13:43:57 +00001098 }
drh6ae49e62021-12-05 20:19:47 +00001099 sqlite3VdbeResolveLabel(v, addrCont);
drh50fb7e02021-12-06 20:16:53 +00001100 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1);
drh6ae49e62021-12-05 20:19:47 +00001101 VdbeCoverage(v);
1102 sqlite3VdbeJumpHere(v, addrTop);
1103 pLoop->wsFlags &= ~WHERE_BLOOMFILTER;
1104 if( OptimizationDisabled(pParse->db, SQLITE_BloomPulldown) ) break;
drhc5860af2021-12-13 18:43:46 +00001105 while( ++iLevel < pWInfo->nLevel ){
drhff55da32022-03-04 20:54:09 +00001106 const SrcItem *pTabItem;
drh6ae49e62021-12-05 20:19:47 +00001107 pLevel = &pWInfo->a[iLevel];
drhff55da32022-03-04 20:54:09 +00001108 pTabItem = &pWInfo->pTabList->a[pLevel->iFrom];
drha76ac882022-04-08 19:20:12 +00001109 if( pTabItem->fg.jointype & (JT_LEFT|JT_LTORJ) ) continue;
drh6ae49e62021-12-05 20:19:47 +00001110 pLoop = pLevel->pWLoop;
drh4f2006d2021-12-13 18:53:10 +00001111 if( NEVER(pLoop==0) ) continue;
drh761d64b2021-12-07 22:37:50 +00001112 if( pLoop->prereq & notReady ) continue;
drhc5860af2021-12-13 18:43:46 +00001113 if( (pLoop->wsFlags & (WHERE_BLOOMFILTER|WHERE_COLUMN_IN))
1114 ==WHERE_BLOOMFILTER
1115 ){
drhdc56dc92021-12-11 17:10:58 +00001116 /* This is a candidate for bloom-filter pull-down (early evaluation).
drhc5860af2021-12-13 18:43:46 +00001117 ** The test that WHERE_COLUMN_IN is omitted is important, as we are
1118 ** not able to do early evaluation of bloom filters that make use of
1119 ** the IN operator */
drhdc56dc92021-12-11 17:10:58 +00001120 break;
1121 }
drh6ae49e62021-12-05 20:19:47 +00001122 }
1123 }while( iLevel < pWInfo->nLevel );
1124 sqlite3VdbeJumpHere(v, addrOnce);
drhfa35f5c2021-12-04 13:43:57 +00001125}
1126
1127
drh9eff6162006-06-12 21:59:13 +00001128#ifndef SQLITE_OMIT_VIRTUALTABLE
1129/*
danielk19771d461462009-04-21 09:02:45 +00001130** Allocate and populate an sqlite3_index_info structure. It is the
1131** responsibility of the caller to eventually release the structure
drh82801a52022-01-20 17:10:59 +00001132** by passing the pointer returned by this function to freeIndexInfo().
danielk19771d461462009-04-21 09:02:45 +00001133*/
drh5346e952013-05-08 14:14:26 +00001134static sqlite3_index_info *allocateIndexInfo(
drhec778d22022-01-22 00:18:01 +00001135 WhereInfo *pWInfo, /* The WHERE clause */
drhefc88d02017-12-22 00:52:50 +00001136 WhereClause *pWC, /* The WHERE clause being analyzed */
dan4f20cd42015-06-08 18:05:54 +00001137 Bitmask mUnusable, /* Ignore terms with these prereqs */
drh76012942021-02-21 21:04:54 +00001138 SrcItem *pSrc, /* The FROM clause term that is the vtab */
dan6256c1c2016-08-08 20:15:41 +00001139 u16 *pmNoOmit /* Mask of terms not to omit */
drh5346e952013-05-08 14:14:26 +00001140){
danielk19771d461462009-04-21 09:02:45 +00001141 int i, j;
1142 int nTerm;
drhec778d22022-01-22 00:18:01 +00001143 Parse *pParse = pWInfo->pParse;
danielk19771d461462009-04-21 09:02:45 +00001144 struct sqlite3_index_constraint *pIdxCons;
1145 struct sqlite3_index_orderby *pIdxOrderBy;
1146 struct sqlite3_index_constraint_usage *pUsage;
drhefc88d02017-12-22 00:52:50 +00001147 struct HiddenIndexInfo *pHidden;
danielk19771d461462009-04-21 09:02:45 +00001148 WhereTerm *pTerm;
1149 int nOrderBy;
1150 sqlite3_index_info *pIdxInfo;
dan6256c1c2016-08-08 20:15:41 +00001151 u16 mNoOmit = 0;
drh8a95d3d2021-12-15 20:48:15 +00001152 const Table *pTab;
drhec778d22022-01-22 00:18:01 +00001153 int eDistinct = 0;
1154 ExprList *pOrderBy = pWInfo->pOrderBy;
1155
drh52576b72021-12-14 20:13:28 +00001156 assert( pSrc!=0 );
drh8a95d3d2021-12-15 20:48:15 +00001157 pTab = pSrc->pTab;
1158 assert( pTab!=0 );
1159 assert( IsVirtual(pTab) );
drh52576b72021-12-14 20:13:28 +00001160
drh8a95d3d2021-12-15 20:48:15 +00001161 /* Find all WHERE clause constraints referring to this virtual table.
1162 ** Mark each term with the TERM_OK flag. Set nTerm to the number of
1163 ** terms found.
1164 */
danielk19771d461462009-04-21 09:02:45 +00001165 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
drh8a95d3d2021-12-15 20:48:15 +00001166 pTerm->wtFlags &= ~TERM_OK;
danielk19771d461462009-04-21 09:02:45 +00001167 if( pTerm->leftCursor != pSrc->iCursor ) continue;
dan4f20cd42015-06-08 18:05:54 +00001168 if( pTerm->prereqRight & mUnusable ) continue;
drh7a5bcc02013-01-16 17:08:58 +00001169 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) );
1170 testcase( pTerm->eOperator & WO_IN );
1171 testcase( pTerm->eOperator & WO_ISNULL );
drhee145872015-05-14 13:18:47 +00001172 testcase( pTerm->eOperator & WO_IS );
dana4ff8252014-01-20 19:55:33 +00001173 testcase( pTerm->eOperator & WO_ALL );
dand03024d2017-09-09 19:41:12 +00001174 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue;
drhb4256992011-08-02 01:57:39 +00001175 if( pTerm->wtFlags & TERM_VNULL ) continue;
drh0fe7e7d2022-02-01 14:58:29 +00001176
drh220f0d62021-10-15 17:06:16 +00001177 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 );
drh8a95d3d2021-12-15 20:48:15 +00001178 assert( pTerm->u.x.leftColumn>=XN_ROWID );
1179 assert( pTerm->u.x.leftColumn<pTab->nCol );
1180
1181 /* tag-20191211-002: WHERE-clause constraints are not useful to the
drha76ac882022-04-08 19:20:12 +00001182 ** right-hand table of a LEFT JOIN nor to the left-hand table of a
1183 ** RIGHT JOIN. See tag-20191211-001 for the
drh8a95d3d2021-12-15 20:48:15 +00001184 ** equivalent restriction for ordinary tables. */
drha76ac882022-04-08 19:20:12 +00001185 if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ))!=0
drh8a95d3d2021-12-15 20:48:15 +00001186 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin)
1187 ){
1188 continue;
1189 }
danielk19771d461462009-04-21 09:02:45 +00001190 nTerm++;
drh8a95d3d2021-12-15 20:48:15 +00001191 pTerm->wtFlags |= TERM_OK;
danielk19771d461462009-04-21 09:02:45 +00001192 }
1193
1194 /* If the ORDER BY clause contains only columns in the current
1195 ** virtual table then allocate space for the aOrderBy part of
1196 ** the sqlite3_index_info structure.
1197 */
1198 nOrderBy = 0;
1199 if( pOrderBy ){
drh56f1b992012-09-25 14:29:39 +00001200 int n = pOrderBy->nExpr;
1201 for(i=0; i<n; i++){
danielk19771d461462009-04-21 09:02:45 +00001202 Expr *pExpr = pOrderBy->a[i].pExpr;
drh52576b72021-12-14 20:13:28 +00001203 Expr *pE2;
1204
drhe1961c52021-12-30 17:36:54 +00001205 /* Skip over constant terms in the ORDER BY clause */
1206 if( sqlite3ExprIsConstant(pExpr) ){
1207 continue;
1208 }
1209
drh52576b72021-12-14 20:13:28 +00001210 /* Virtual tables are unable to deal with NULLS FIRST */
dan4fcb9ca2019-08-20 15:47:28 +00001211 if( pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL ) break;
drh52576b72021-12-14 20:13:28 +00001212
1213 /* First case - a direct column references without a COLLATE operator */
1214 if( pExpr->op==TK_COLUMN && pExpr->iTable==pSrc->iCursor ){
drh8a95d3d2021-12-15 20:48:15 +00001215 assert( pExpr->iColumn>=XN_ROWID && pExpr->iColumn<pTab->nCol );
drh52576b72021-12-14 20:13:28 +00001216 continue;
1217 }
1218
1219 /* 2nd case - a column reference with a COLLATE operator. Only match
1220 ** of the COLLATE operator matches the collation of the column. */
1221 if( pExpr->op==TK_COLLATE
1222 && (pE2 = pExpr->pLeft)->op==TK_COLUMN
1223 && pE2->iTable==pSrc->iCursor
1224 ){
1225 const char *zColl; /* The collating sequence name */
1226 assert( !ExprHasProperty(pExpr, EP_IntValue) );
1227 assert( pExpr->u.zToken!=0 );
drh8a95d3d2021-12-15 20:48:15 +00001228 assert( pE2->iColumn>=XN_ROWID && pE2->iColumn<pTab->nCol );
drh52576b72021-12-14 20:13:28 +00001229 pExpr->iColumn = pE2->iColumn;
1230 if( pE2->iColumn<0 ) continue; /* Collseq does not matter for rowid */
drh8a95d3d2021-12-15 20:48:15 +00001231 zColl = sqlite3ColumnColl(&pTab->aCol[pE2->iColumn]);
drh52576b72021-12-14 20:13:28 +00001232 if( zColl==0 ) zColl = sqlite3StrBINARY;
1233 if( sqlite3_stricmp(pExpr->u.zToken, zColl)==0 ) continue;
1234 }
1235
1236 /* No matches cause a break out of the loop */
1237 break;
danielk19771d461462009-04-21 09:02:45 +00001238 }
drh0fe7e7d2022-02-01 14:58:29 +00001239 if( i==n ){
drh56f1b992012-09-25 14:29:39 +00001240 nOrderBy = n;
danece092e2022-03-16 12:06:00 +00001241 if( (pWInfo->wctrlFlags & WHERE_DISTINCTBY) ){
1242 eDistinct = 2 + ((pWInfo->wctrlFlags & WHERE_SORTBYGROUP)!=0);
1243 }else if( pWInfo->wctrlFlags & WHERE_GROUPBY ){
1244 eDistinct = 1;
drhec778d22022-01-22 00:18:01 +00001245 }
danielk19771d461462009-04-21 09:02:45 +00001246 }
1247 }
1248
1249 /* Allocate the sqlite3_index_info structure
1250 */
1251 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo)
1252 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm
drh82801a52022-01-20 17:10:59 +00001253 + sizeof(*pIdxOrderBy)*nOrderBy + sizeof(*pHidden)
1254 + sizeof(sqlite3_value*)*nTerm );
danielk19771d461462009-04-21 09:02:45 +00001255 if( pIdxInfo==0 ){
1256 sqlite3ErrorMsg(pParse, "out of memory");
danielk19771d461462009-04-21 09:02:45 +00001257 return 0;
1258 }
drhefc88d02017-12-22 00:52:50 +00001259 pHidden = (struct HiddenIndexInfo*)&pIdxInfo[1];
drh82801a52022-01-20 17:10:59 +00001260 pIdxCons = (struct sqlite3_index_constraint*)&pHidden->aRhs[nTerm];
danielk19771d461462009-04-21 09:02:45 +00001261 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm];
1262 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy];
drhcfcf4de2019-12-28 13:01:52 +00001263 pIdxInfo->aConstraint = pIdxCons;
1264 pIdxInfo->aOrderBy = pIdxOrderBy;
1265 pIdxInfo->aConstraintUsage = pUsage;
drhefc88d02017-12-22 00:52:50 +00001266 pHidden->pWC = pWC;
1267 pHidden->pParse = pParse;
drhec778d22022-01-22 00:18:01 +00001268 pHidden->eDistinct = eDistinct;
drha9f18f02022-02-01 16:30:57 +00001269 pHidden->mIn = 0;
danielk19771d461462009-04-21 09:02:45 +00001270 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
dand03024d2017-09-09 19:41:12 +00001271 u16 op;
drh8a95d3d2021-12-15 20:48:15 +00001272 if( (pTerm->wtFlags & TERM_OK)==0 ) continue;
drh75fa2662020-09-28 15:49:43 +00001273 pIdxCons[j].iColumn = pTerm->u.x.leftColumn;
danielk19771d461462009-04-21 09:02:45 +00001274 pIdxCons[j].iTermOffset = i;
dand03024d2017-09-09 19:41:12 +00001275 op = pTerm->eOperator & WO_ALL;
drha9f18f02022-02-01 16:30:57 +00001276 if( op==WO_IN ){
drh51896e62022-02-06 22:13:35 +00001277 if( (pTerm->wtFlags & TERM_SLICE)==0 ){
1278 pHidden->mIn |= SMASKBIT32(j);
1279 }
drha9f18f02022-02-01 16:30:57 +00001280 op = WO_EQ;
1281 }
drh303a69b2017-09-11 19:47:37 +00001282 if( op==WO_AUX ){
dand03024d2017-09-09 19:41:12 +00001283 pIdxCons[j].op = pTerm->eMatchOp;
1284 }else if( op & (WO_ISNULL|WO_IS) ){
1285 if( op==WO_ISNULL ){
1286 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_ISNULL;
1287 }else{
1288 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_IS;
1289 }
1290 }else{
1291 pIdxCons[j].op = (u8)op;
1292 /* The direct assignment in the previous line is possible only because
1293 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The
1294 ** following asserts verify this fact. */
1295 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ );
1296 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT );
1297 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE );
1298 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT );
1299 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE );
drh303a69b2017-09-11 19:47:37 +00001300 assert( pTerm->eOperator&(WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_AUX) );
dan6256c1c2016-08-08 20:15:41 +00001301
dand03024d2017-09-09 19:41:12 +00001302 if( op & (WO_LT|WO_LE|WO_GT|WO_GE)
drh33892c12017-09-11 18:37:44 +00001303 && sqlite3ExprIsVector(pTerm->pExpr->pRight)
1304 ){
drhb6c94722019-12-05 21:46:23 +00001305 testcase( j!=i );
1306 if( j<16 ) mNoOmit |= (1 << j);
dand03024d2017-09-09 19:41:12 +00001307 if( op==WO_LT ) pIdxCons[j].op = WO_LE;
1308 if( op==WO_GT ) pIdxCons[j].op = WO_GE;
1309 }
dan6256c1c2016-08-08 20:15:41 +00001310 }
1311
danielk19771d461462009-04-21 09:02:45 +00001312 j++;
1313 }
drh8a95d3d2021-12-15 20:48:15 +00001314 assert( j==nTerm );
drhcfcf4de2019-12-28 13:01:52 +00001315 pIdxInfo->nConstraint = j;
drhe1961c52021-12-30 17:36:54 +00001316 for(i=j=0; i<nOrderBy; i++){
danielk19771d461462009-04-21 09:02:45 +00001317 Expr *pExpr = pOrderBy->a[i].pExpr;
drhe1961c52021-12-30 17:36:54 +00001318 if( sqlite3ExprIsConstant(pExpr) ) continue;
drh52576b72021-12-14 20:13:28 +00001319 assert( pExpr->op==TK_COLUMN
1320 || (pExpr->op==TK_COLLATE && pExpr->pLeft->op==TK_COLUMN
1321 && pExpr->iColumn==pExpr->pLeft->iColumn) );
drhe1961c52021-12-30 17:36:54 +00001322 pIdxOrderBy[j].iColumn = pExpr->iColumn;
1323 pIdxOrderBy[j].desc = pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC;
1324 j++;
danielk19771d461462009-04-21 09:02:45 +00001325 }
drhe1961c52021-12-30 17:36:54 +00001326 pIdxInfo->nOrderBy = j;
danielk19771d461462009-04-21 09:02:45 +00001327
dan6256c1c2016-08-08 20:15:41 +00001328 *pmNoOmit = mNoOmit;
danielk19771d461462009-04-21 09:02:45 +00001329 return pIdxInfo;
1330}
1331
1332/*
drh82801a52022-01-20 17:10:59 +00001333** Free an sqlite3_index_info structure allocated by allocateIndexInfo()
1334** and possibly modified by xBestIndex methods.
1335*/
1336static void freeIndexInfo(sqlite3 *db, sqlite3_index_info *pIdxInfo){
1337 HiddenIndexInfo *pHidden;
1338 int i;
1339 assert( pIdxInfo!=0 );
1340 pHidden = (HiddenIndexInfo*)&pIdxInfo[1];
1341 assert( pHidden->pParse!=0 );
1342 assert( pHidden->pParse->db==db );
1343 for(i=0; i<pIdxInfo->nConstraint; i++){
drh991d1082022-01-21 00:38:49 +00001344 sqlite3ValueFree(pHidden->aRhs[i]); /* IMP: R-14553-25174 */
drh82801a52022-01-20 17:10:59 +00001345 pHidden->aRhs[i] = 0;
1346 }
1347 sqlite3DbFree(db, pIdxInfo);
1348}
1349
1350/*
danielk19771d461462009-04-21 09:02:45 +00001351** The table object reference passed as the second argument to this function
1352** must represent a virtual table. This function invokes the xBestIndex()
drh3b48e8c2013-06-12 20:18:16 +00001353** method of the virtual table with the sqlite3_index_info object that
1354** comes in as the 3rd argument to this function.
danielk19771d461462009-04-21 09:02:45 +00001355**
drh32dcc842018-11-16 13:56:15 +00001356** If an error occurs, pParse is populated with an error message and an
1357** appropriate error code is returned. A return of SQLITE_CONSTRAINT from
1358** xBestIndex is not considered an error. SQLITE_CONSTRAINT indicates that
1359** the current configuration of "unusable" flags in sqlite3_index_info can
1360** not result in a valid plan.
danielk19771d461462009-04-21 09:02:45 +00001361**
1362** Whether or not an error is returned, it is the responsibility of the
1363** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates
1364** that this is required.
1365*/
1366static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){
danielk1977595a5232009-07-24 17:58:53 +00001367 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab;
danielk19771d461462009-04-21 09:02:45 +00001368 int rc;
1369
drhcfcf4de2019-12-28 13:01:52 +00001370 whereTraceIndexInfoInputs(p);
drh3832c3c2022-02-03 14:19:26 +00001371 pParse->db->nSchemaLock++;
danielk19771d461462009-04-21 09:02:45 +00001372 rc = pVtab->pModule->xBestIndex(pVtab, p);
drh3832c3c2022-02-03 14:19:26 +00001373 pParse->db->nSchemaLock--;
drhcfcf4de2019-12-28 13:01:52 +00001374 whereTraceIndexInfoOutputs(p);
danielk19771d461462009-04-21 09:02:45 +00001375
drh32dcc842018-11-16 13:56:15 +00001376 if( rc!=SQLITE_OK && rc!=SQLITE_CONSTRAINT ){
danielk19771d461462009-04-21 09:02:45 +00001377 if( rc==SQLITE_NOMEM ){
drh4a642b62016-02-05 01:55:27 +00001378 sqlite3OomFault(pParse->db);
danielk19771d461462009-04-21 09:02:45 +00001379 }else if( !pVtab->zErrMsg ){
1380 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc));
1381 }else{
1382 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg);
1383 }
1384 }
drhb9755982010-07-24 16:34:37 +00001385 sqlite3_free(pVtab->zErrMsg);
danielk19771d461462009-04-21 09:02:45 +00001386 pVtab->zErrMsg = 0;
drh32dcc842018-11-16 13:56:15 +00001387 return rc;
danielk19771d461462009-04-21 09:02:45 +00001388}
drh7ba39a92013-05-30 17:43:19 +00001389#endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */
danielk19771d461462009-04-21 09:02:45 +00001390
drh175b8f02019-08-08 15:24:17 +00001391#ifdef SQLITE_ENABLE_STAT4
drh28c4cf42005-07-27 20:41:43 +00001392/*
drhfaacf172011-08-12 01:51:45 +00001393** Estimate the location of a particular key among all keys in an
1394** index. Store the results in aStat as follows:
drhe847d322011-01-20 02:56:37 +00001395**
dana3d0c132015-03-14 18:59:58 +00001396** aStat[0] Est. number of rows less than pRec
1397** aStat[1] Est. number of rows equal to pRec
dan02fa4692009-08-17 17:06:58 +00001398**
drh6d3f91d2014-11-05 19:26:12 +00001399** Return the index of the sample that is the smallest sample that
dana3d0c132015-03-14 18:59:58 +00001400** is greater than or equal to pRec. Note that this index is not an index
1401** into the aSample[] array - it is an index into a virtual set of samples
1402** based on the contents of aSample[] and the number of fields in record
1403** pRec.
dan02fa4692009-08-17 17:06:58 +00001404*/
drh6d3f91d2014-11-05 19:26:12 +00001405static int whereKeyStats(
dan02fa4692009-08-17 17:06:58 +00001406 Parse *pParse, /* Database connection */
1407 Index *pIdx, /* Index to consider domain of */
dan7a419232013-08-06 20:01:43 +00001408 UnpackedRecord *pRec, /* Vector of values to consider */
drhfaacf172011-08-12 01:51:45 +00001409 int roundUp, /* Round up if true. Round down if false */
1410 tRowcnt *aStat /* OUT: stats written here */
dan02fa4692009-08-17 17:06:58 +00001411){
danf52bb8d2013-08-03 20:24:58 +00001412 IndexSample *aSample = pIdx->aSample;
drhfbc38de2013-09-03 19:26:22 +00001413 int iCol; /* Index of required stats in anEq[] etc. */
dana3d0c132015-03-14 18:59:58 +00001414 int i; /* Index of first sample >= pRec */
1415 int iSample; /* Smallest sample larger than or equal to pRec */
dan84c309b2013-08-08 16:17:12 +00001416 int iMin = 0; /* Smallest sample not yet tested */
dan84c309b2013-08-08 16:17:12 +00001417 int iTest; /* Next sample to test */
1418 int res; /* Result of comparison operation */
dana3d0c132015-03-14 18:59:58 +00001419 int nField; /* Number of fields in pRec */
1420 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */
dan02fa4692009-08-17 17:06:58 +00001421
drh4f991892013-10-11 15:05:05 +00001422#ifndef SQLITE_DEBUG
1423 UNUSED_PARAMETER( pParse );
1424#endif
drh7f594752013-12-03 19:49:55 +00001425 assert( pRec!=0 );
drh5c624862011-09-22 18:46:34 +00001426 assert( pIdx->nSample>0 );
dana3d0c132015-03-14 18:59:58 +00001427 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol );
1428
1429 /* Do a binary search to find the first sample greater than or equal
1430 ** to pRec. If pRec contains a single field, the set of samples to search
1431 ** is simply the aSample[] array. If the samples in aSample[] contain more
1432 ** than one fields, all fields following the first are ignored.
1433 **
1434 ** If pRec contains N fields, where N is more than one, then as well as the
1435 ** samples in aSample[] (truncated to N fields), the search also has to
1436 ** consider prefixes of those samples. For example, if the set of samples
1437 ** in aSample is:
1438 **
1439 ** aSample[0] = (a, 5)
1440 ** aSample[1] = (a, 10)
1441 ** aSample[2] = (b, 5)
1442 ** aSample[3] = (c, 100)
1443 ** aSample[4] = (c, 105)
1444 **
1445 ** Then the search space should ideally be the samples above and the
1446 ** unique prefixes [a], [b] and [c]. But since that is hard to organize,
1447 ** the code actually searches this set:
1448 **
1449 ** 0: (a)
1450 ** 1: (a, 5)
1451 ** 2: (a, 10)
1452 ** 3: (a, 10)
1453 ** 4: (b)
1454 ** 5: (b, 5)
1455 ** 6: (c)
1456 ** 7: (c, 100)
1457 ** 8: (c, 105)
1458 ** 9: (c, 105)
1459 **
1460 ** For each sample in the aSample[] array, N samples are present in the
1461 ** effective sample array. In the above, samples 0 and 1 are based on
1462 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc.
1463 **
1464 ** Often, sample i of each block of N effective samples has (i+1) fields.
1465 ** Except, each sample may be extended to ensure that it is greater than or
1466 ** equal to the previous sample in the array. For example, in the above,
1467 ** sample 2 is the first sample of a block of N samples, so at first it
1468 ** appears that it should be 1 field in size. However, that would make it
1469 ** smaller than sample 1, so the binary search would not work. As a result,
1470 ** it is extended to two fields. The duplicates that this creates do not
1471 ** cause any problems.
1472 */
1473 nField = pRec->nField;
1474 iCol = 0;
1475 iSample = pIdx->nSample * nField;
dan84c309b2013-08-08 16:17:12 +00001476 do{
dana3d0c132015-03-14 18:59:58 +00001477 int iSamp; /* Index in aSample[] of test sample */
1478 int n; /* Number of fields in test sample */
1479
1480 iTest = (iMin+iSample)/2;
1481 iSamp = iTest / nField;
1482 if( iSamp>0 ){
1483 /* The proposed effective sample is a prefix of sample aSample[iSamp].
1484 ** Specifically, the shortest prefix of at least (1 + iTest%nField)
1485 ** fields that is greater than the previous effective sample. */
1486 for(n=(iTest % nField) + 1; n<nField; n++){
1487 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break;
1488 }
dan84c309b2013-08-08 16:17:12 +00001489 }else{
dana3d0c132015-03-14 18:59:58 +00001490 n = iTest + 1;
dan02fa4692009-08-17 17:06:58 +00001491 }
dana3d0c132015-03-14 18:59:58 +00001492
1493 pRec->nField = n;
1494 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec);
1495 if( res<0 ){
1496 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1];
1497 iMin = iTest+1;
1498 }else if( res==0 && n<nField ){
1499 iLower = aSample[iSamp].anLt[n-1];
1500 iMin = iTest+1;
1501 res = -1;
1502 }else{
1503 iSample = iTest;
1504 iCol = n-1;
1505 }
1506 }while( res && iMin<iSample );
1507 i = iSample / nField;
drh51147ba2005-07-23 22:59:55 +00001508
dan84c309b2013-08-08 16:17:12 +00001509#ifdef SQLITE_DEBUG
1510 /* The following assert statements check that the binary search code
1511 ** above found the right answer. This block serves no purpose other
1512 ** than to invoke the asserts. */
dana3d0c132015-03-14 18:59:58 +00001513 if( pParse->db->mallocFailed==0 ){
1514 if( res==0 ){
1515 /* If (res==0) is true, then pRec must be equal to sample i. */
1516 assert( i<pIdx->nSample );
1517 assert( iCol==nField-1 );
1518 pRec->nField = nField;
1519 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)
1520 || pParse->db->mallocFailed
1521 );
1522 }else{
1523 /* Unless i==pIdx->nSample, indicating that pRec is larger than
1524 ** all samples in the aSample[] array, pRec must be smaller than the
1525 ** (iCol+1) field prefix of sample i. */
1526 assert( i<=pIdx->nSample && i>=0 );
1527 pRec->nField = iCol+1;
1528 assert( i==pIdx->nSample
1529 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0
1530 || pParse->db->mallocFailed );
1531
1532 /* if i==0 and iCol==0, then record pRec is smaller than all samples
1533 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must
1534 ** be greater than or equal to the (iCol) field prefix of sample i.
1535 ** If (i>0), then pRec must also be greater than sample (i-1). */
1536 if( iCol>0 ){
1537 pRec->nField = iCol;
1538 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0
1539 || pParse->db->mallocFailed );
1540 }
1541 if( i>0 ){
1542 pRec->nField = nField;
1543 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0
1544 || pParse->db->mallocFailed );
1545 }
1546 }
drhfaacf172011-08-12 01:51:45 +00001547 }
dan84c309b2013-08-08 16:17:12 +00001548#endif /* ifdef SQLITE_DEBUG */
dan02fa4692009-08-17 17:06:58 +00001549
dan84c309b2013-08-08 16:17:12 +00001550 if( res==0 ){
dana3d0c132015-03-14 18:59:58 +00001551 /* Record pRec is equal to sample i */
1552 assert( iCol==nField-1 );
daneea568d2013-08-07 19:46:15 +00001553 aStat[0] = aSample[i].anLt[iCol];
1554 aStat[1] = aSample[i].anEq[iCol];
drhfaacf172011-08-12 01:51:45 +00001555 }else{
dana3d0c132015-03-14 18:59:58 +00001556 /* At this point, the (iCol+1) field prefix of aSample[i] is the first
1557 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec
1558 ** is larger than all samples in the array. */
1559 tRowcnt iUpper, iGap;
1560 if( i>=pIdx->nSample ){
1561 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]);
drhfaacf172011-08-12 01:51:45 +00001562 }else{
dana3d0c132015-03-14 18:59:58 +00001563 iUpper = aSample[i].anLt[iCol];
drhfaacf172011-08-12 01:51:45 +00001564 }
dana3d0c132015-03-14 18:59:58 +00001565
drhfaacf172011-08-12 01:51:45 +00001566 if( iLower>=iUpper ){
1567 iGap = 0;
1568 }else{
1569 iGap = iUpper - iLower;
drhfaacf172011-08-12 01:51:45 +00001570 }
1571 if( roundUp ){
1572 iGap = (iGap*2)/3;
1573 }else{
1574 iGap = iGap/3;
1575 }
1576 aStat[0] = iLower + iGap;
drh63ad86e2017-05-24 04:18:00 +00001577 aStat[1] = pIdx->aAvgEq[nField-1];
dan02fa4692009-08-17 17:06:58 +00001578 }
dana3d0c132015-03-14 18:59:58 +00001579
1580 /* Restore the pRec->nField value before returning. */
1581 pRec->nField = nField;
drh6d3f91d2014-11-05 19:26:12 +00001582 return i;
dan02fa4692009-08-17 17:06:58 +00001583}
drh175b8f02019-08-08 15:24:17 +00001584#endif /* SQLITE_ENABLE_STAT4 */
dan937d0de2009-10-15 18:35:38 +00001585
1586/*
danaa9933c2014-04-24 20:04:49 +00001587** If it is not NULL, pTerm is a term that provides an upper or lower
1588** bound on a range scan. Without considering pTerm, it is estimated
1589** that the scan will visit nNew rows. This function returns the number
1590** estimated to be visited after taking pTerm into account.
1591**
1592** If the user explicitly specified a likelihood() value for this term,
1593** then the return value is the likelihood multiplied by the number of
1594** input rows. Otherwise, this function assumes that an "IS NOT NULL" term
1595** has a likelihood of 0.50, and any other term a likelihood of 0.25.
1596*/
1597static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){
1598 LogEst nRet = nNew;
1599 if( pTerm ){
1600 if( pTerm->truthProb<=0 ){
1601 nRet += pTerm->truthProb;
dan7de2a1f2014-04-28 20:11:20 +00001602 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){
danaa9933c2014-04-24 20:04:49 +00001603 nRet -= 20; assert( 20==sqlite3LogEst(4) );
1604 }
1605 }
1606 return nRet;
1607}
1608
drh567cc1e2015-08-25 19:42:28 +00001609
drh175b8f02019-08-08 15:24:17 +00001610#ifdef SQLITE_ENABLE_STAT4
drh567cc1e2015-08-25 19:42:28 +00001611/*
1612** Return the affinity for a single column of an index.
1613*/
dand66e5792016-08-03 16:14:33 +00001614char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){
drh8ffddeb2015-09-25 01:09:27 +00001615 assert( iCol>=0 && iCol<pIdx->nColumn );
drh567cc1e2015-08-25 19:42:28 +00001616 if( !pIdx->zColAff ){
1617 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB;
1618 }
drh96fb16e2019-08-06 14:37:24 +00001619 assert( pIdx->zColAff[iCol]!=0 );
drh567cc1e2015-08-25 19:42:28 +00001620 return pIdx->zColAff[iCol];
1621}
1622#endif
1623
1624
drh175b8f02019-08-08 15:24:17 +00001625#ifdef SQLITE_ENABLE_STAT4
danb0b82902014-06-26 20:21:46 +00001626/*
1627** This function is called to estimate the number of rows visited by a
1628** range-scan on a skip-scan index. For example:
1629**
1630** CREATE INDEX i1 ON t1(a, b, c);
1631** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?;
1632**
1633** Value pLoop->nOut is currently set to the estimated number of rows
1634** visited for scanning (a=? AND b=?). This function reduces that estimate
1635** by some factor to account for the (c BETWEEN ? AND ?) expression based
1636** on the stat4 data for the index. this scan will be peformed multiple
1637** times (once for each (a,b) combination that matches a=?) is dealt with
1638** by the caller.
1639**
1640** It does this by scanning through all stat4 samples, comparing values
1641** extracted from pLower and pUpper with the corresponding column in each
1642** sample. If L and U are the number of samples found to be less than or
1643** equal to the values extracted from pLower and pUpper respectively, and
1644** N is the total number of samples, the pLoop->nOut value is adjusted
1645** as follows:
1646**
1647** nOut = nOut * ( min(U - L, 1) / N )
1648**
1649** If pLower is NULL, or a value cannot be extracted from the term, L is
1650** set to zero. If pUpper is NULL, or a value cannot be extracted from it,
1651** U is set to N.
1652**
1653** Normally, this function sets *pbDone to 1 before returning. However,
1654** if no value can be extracted from either pLower or pUpper (and so the
1655** estimate of the number of rows delivered remains unchanged), *pbDone
1656** is left as is.
1657**
1658** If an error occurs, an SQLite error code is returned. Otherwise,
1659** SQLITE_OK.
1660*/
1661static int whereRangeSkipScanEst(
1662 Parse *pParse, /* Parsing & code generating context */
1663 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */
1664 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */
1665 WhereLoop *pLoop, /* Update the .nOut value of this loop */
1666 int *pbDone /* Set to true if at least one expr. value extracted */
1667){
1668 Index *p = pLoop->u.btree.pIndex;
1669 int nEq = pLoop->u.btree.nEq;
1670 sqlite3 *db = pParse->db;
dan4e42ba42014-06-27 20:14:25 +00001671 int nLower = -1;
1672 int nUpper = p->nSample+1;
danb0b82902014-06-26 20:21:46 +00001673 int rc = SQLITE_OK;
drh8ffddeb2015-09-25 01:09:27 +00001674 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq);
danb0b82902014-06-26 20:21:46 +00001675 CollSeq *pColl;
1676
1677 sqlite3_value *p1 = 0; /* Value extracted from pLower */
1678 sqlite3_value *p2 = 0; /* Value extracted from pUpper */
1679 sqlite3_value *pVal = 0; /* Value extracted from record */
1680
1681 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]);
1682 if( pLower ){
1683 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1);
dan4e42ba42014-06-27 20:14:25 +00001684 nLower = 0;
danb0b82902014-06-26 20:21:46 +00001685 }
1686 if( pUpper && rc==SQLITE_OK ){
1687 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2);
dan4e42ba42014-06-27 20:14:25 +00001688 nUpper = p2 ? 0 : p->nSample;
danb0b82902014-06-26 20:21:46 +00001689 }
1690
1691 if( p1 || p2 ){
1692 int i;
1693 int nDiff;
1694 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){
1695 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal);
1696 if( rc==SQLITE_OK && p1 ){
1697 int res = sqlite3MemCompare(p1, pVal, pColl);
dan4e42ba42014-06-27 20:14:25 +00001698 if( res>=0 ) nLower++;
danb0b82902014-06-26 20:21:46 +00001699 }
1700 if( rc==SQLITE_OK && p2 ){
1701 int res = sqlite3MemCompare(p2, pVal, pColl);
dan4e42ba42014-06-27 20:14:25 +00001702 if( res>=0 ) nUpper++;
danb0b82902014-06-26 20:21:46 +00001703 }
1704 }
danb0b82902014-06-26 20:21:46 +00001705 nDiff = (nUpper - nLower);
1706 if( nDiff<=0 ) nDiff = 1;
dan4e42ba42014-06-27 20:14:25 +00001707
1708 /* If there is both an upper and lower bound specified, and the
1709 ** comparisons indicate that they are close together, use the fallback
1710 ** method (assume that the scan visits 1/64 of the rows) for estimating
1711 ** the number of rows visited. Otherwise, estimate the number of rows
1712 ** using the method described in the header comment for this function. */
1713 if( nDiff!=1 || pUpper==0 || pLower==0 ){
1714 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff));
1715 pLoop->nOut -= nAdjust;
1716 *pbDone = 1;
1717 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n",
danfa887452014-06-28 15:26:10 +00001718 nLower, nUpper, nAdjust*-1, pLoop->nOut));
dan4e42ba42014-06-27 20:14:25 +00001719 }
1720
danb0b82902014-06-26 20:21:46 +00001721 }else{
1722 assert( *pbDone==0 );
1723 }
1724
1725 sqlite3ValueFree(p1);
1726 sqlite3ValueFree(p2);
1727 sqlite3ValueFree(pVal);
1728
1729 return rc;
1730}
drh175b8f02019-08-08 15:24:17 +00001731#endif /* SQLITE_ENABLE_STAT4 */
danb0b82902014-06-26 20:21:46 +00001732
danaa9933c2014-04-24 20:04:49 +00001733/*
dan02fa4692009-08-17 17:06:58 +00001734** This function is used to estimate the number of rows that will be visited
1735** by scanning an index for a range of values. The range may have an upper
1736** bound, a lower bound, or both. The WHERE clause terms that set the upper
1737** and lower bounds are represented by pLower and pUpper respectively. For
1738** example, assuming that index p is on t1(a):
1739**
1740** ... FROM t1 WHERE a > ? AND a < ? ...
1741** |_____| |_____|
1742** | |
1743** pLower pUpper
1744**
drh98cdf622009-08-20 18:14:42 +00001745** If either of the upper or lower bound is not present, then NULL is passed in
drhcdaca552009-08-20 13:45:07 +00001746** place of the corresponding WhereTerm.
dan02fa4692009-08-17 17:06:58 +00001747**
drh6d3f91d2014-11-05 19:26:12 +00001748** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index
dan6cb8d762013-08-08 11:48:57 +00001749** column subject to the range constraint. Or, equivalently, the number of
1750** equality constraints optimized by the proposed index scan. For example,
1751** assuming index p is on t1(a, b), and the SQL query is:
dan02fa4692009-08-17 17:06:58 +00001752**
1753** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ...
1754**
dan6cb8d762013-08-08 11:48:57 +00001755** then nEq is set to 1 (as the range restricted column, b, is the second
1756** left-most column of the index). Or, if the query is:
dan02fa4692009-08-17 17:06:58 +00001757**
1758** ... FROM t1 WHERE a > ? AND a < ? ...
1759**
dan6cb8d762013-08-08 11:48:57 +00001760** then nEq is set to 0.
dan02fa4692009-08-17 17:06:58 +00001761**
drhbf539c42013-10-05 18:16:02 +00001762** When this function is called, *pnOut is set to the sqlite3LogEst() of the
dan6cb8d762013-08-08 11:48:57 +00001763** number of rows that the index scan is expected to visit without
drh6d3f91d2014-11-05 19:26:12 +00001764** considering the range constraints. If nEq is 0, then *pnOut is the number of
dan6cb8d762013-08-08 11:48:57 +00001765** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced)
peter.d.reid60ec9142014-09-06 16:39:46 +00001766** to account for the range constraints pLower and pUpper.
dan6cb8d762013-08-08 11:48:57 +00001767**
1768** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be
drh94aa7e02014-06-06 17:09:52 +00001769** used, a single range inequality reduces the search space by a factor of 4.
1770** and a pair of constraints (x>? AND x<?) reduces the expected number of
1771** rows visited by a factor of 64.
dan02fa4692009-08-17 17:06:58 +00001772*/
1773static int whereRangeScanEst(
drhcdaca552009-08-20 13:45:07 +00001774 Parse *pParse, /* Parsing & code generating context */
dan7a419232013-08-06 20:01:43 +00001775 WhereLoopBuilder *pBuilder,
drhcdaca552009-08-20 13:45:07 +00001776 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */
1777 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */
drh186ad8c2013-10-08 18:40:37 +00001778 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */
dan02fa4692009-08-17 17:06:58 +00001779){
dan69188d92009-08-19 08:18:32 +00001780 int rc = SQLITE_OK;
drh186ad8c2013-10-08 18:40:37 +00001781 int nOut = pLoop->nOut;
drhbf539c42013-10-05 18:16:02 +00001782 LogEst nNew;
dan69188d92009-08-19 08:18:32 +00001783
drh175b8f02019-08-08 15:24:17 +00001784#ifdef SQLITE_ENABLE_STAT4
drh186ad8c2013-10-08 18:40:37 +00001785 Index *p = pLoop->u.btree.pIndex;
drh4f991892013-10-11 15:05:05 +00001786 int nEq = pLoop->u.btree.nEq;
dan02fa4692009-08-17 17:06:58 +00001787
drh5eae1d12019-08-08 16:23:12 +00001788 if( p->nSample>0 && ALWAYS(nEq<p->nSampleCol)
1789 && OptimizationEnabled(pParse->db, SQLITE_Stat4)
drh72d03a62018-07-20 19:24:02 +00001790 ){
danb0b82902014-06-26 20:21:46 +00001791 if( nEq==pBuilder->nRecValid ){
1792 UnpackedRecord *pRec = pBuilder->pRec;
1793 tRowcnt a[2];
dand66e5792016-08-03 16:14:33 +00001794 int nBtm = pLoop->u.btree.nBtm;
1795 int nTop = pLoop->u.btree.nTop;
drh98cdf622009-08-20 18:14:42 +00001796
danb0b82902014-06-26 20:21:46 +00001797 /* Variable iLower will be set to the estimate of the number of rows in
1798 ** the index that are less than the lower bound of the range query. The
1799 ** lower bound being the concatenation of $P and $L, where $P is the
1800 ** key-prefix formed by the nEq values matched against the nEq left-most
1801 ** columns of the index, and $L is the value in pLower.
1802 **
1803 ** Or, if pLower is NULL or $L cannot be extracted from it (because it
1804 ** is not a simple variable or literal value), the lower bound of the
1805 ** range is $P. Due to a quirk in the way whereKeyStats() works, even
1806 ** if $L is available, whereKeyStats() is called for both ($P) and
drh6d3f91d2014-11-05 19:26:12 +00001807 ** ($P:$L) and the larger of the two returned values is used.
danb0b82902014-06-26 20:21:46 +00001808 **
1809 ** Similarly, iUpper is to be set to the estimate of the number of rows
1810 ** less than the upper bound of the range query. Where the upper bound
1811 ** is either ($P) or ($P:$U). Again, even if $U is available, both values
1812 ** of iUpper are requested of whereKeyStats() and the smaller used.
drh6d3f91d2014-11-05 19:26:12 +00001813 **
1814 ** The number of rows between the two bounds is then just iUpper-iLower.
danb0b82902014-06-26 20:21:46 +00001815 */
drh6d3f91d2014-11-05 19:26:12 +00001816 tRowcnt iLower; /* Rows less than the lower bound */
1817 tRowcnt iUpper; /* Rows less than the upper bound */
1818 int iLwrIdx = -2; /* aSample[] for the lower bound */
1819 int iUprIdx = -1; /* aSample[] for the upper bound */
danb3c02e22013-08-08 19:38:40 +00001820
drhb34fc5b2014-08-28 17:20:37 +00001821 if( pRec ){
1822 testcase( pRec->nField!=pBuilder->nRecValid );
1823 pRec->nField = pBuilder->nRecValid;
1824 }
danb0b82902014-06-26 20:21:46 +00001825 /* Determine iLower and iUpper using ($P) only. */
1826 if( nEq==0 ){
1827 iLower = 0;
drh9f07cf72014-10-22 15:27:05 +00001828 iUpper = p->nRowEst0;
danb0b82902014-06-26 20:21:46 +00001829 }else{
1830 /* Note: this call could be optimized away - since the same values must
1831 ** have been requested when testing key $P in whereEqualScanEst(). */
1832 whereKeyStats(pParse, p, pRec, 0, a);
1833 iLower = a[0];
1834 iUpper = a[0] + a[1];
dan6cb8d762013-08-08 11:48:57 +00001835 }
danb0b82902014-06-26 20:21:46 +00001836
drh69afd992014-10-08 02:53:25 +00001837 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 );
1838 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 );
drh681fca02014-10-10 15:01:46 +00001839 assert( p->aSortOrder!=0 );
1840 if( p->aSortOrder[nEq] ){
drh69afd992014-10-08 02:53:25 +00001841 /* The roles of pLower and pUpper are swapped for a DESC index */
1842 SWAP(WhereTerm*, pLower, pUpper);
dand66e5792016-08-03 16:14:33 +00001843 SWAP(int, nBtm, nTop);
drh69afd992014-10-08 02:53:25 +00001844 }
1845
danb0b82902014-06-26 20:21:46 +00001846 /* If possible, improve on the iLower estimate using ($P:$L). */
1847 if( pLower ){
dand66e5792016-08-03 16:14:33 +00001848 int n; /* Values extracted from pExpr */
danb0b82902014-06-26 20:21:46 +00001849 Expr *pExpr = pLower->pExpr->pRight;
dand66e5792016-08-03 16:14:33 +00001850 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n);
1851 if( rc==SQLITE_OK && n ){
danb0b82902014-06-26 20:21:46 +00001852 tRowcnt iNew;
dand66e5792016-08-03 16:14:33 +00001853 u16 mask = WO_GT|WO_LE;
1854 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT);
drh6d3f91d2014-11-05 19:26:12 +00001855 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a);
dand66e5792016-08-03 16:14:33 +00001856 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0);
danb0b82902014-06-26 20:21:46 +00001857 if( iNew>iLower ) iLower = iNew;
1858 nOut--;
danf741e042014-08-25 18:29:38 +00001859 pLower = 0;
danb0b82902014-06-26 20:21:46 +00001860 }
1861 }
1862
1863 /* If possible, improve on the iUpper estimate using ($P:$U). */
1864 if( pUpper ){
dand66e5792016-08-03 16:14:33 +00001865 int n; /* Values extracted from pExpr */
danb0b82902014-06-26 20:21:46 +00001866 Expr *pExpr = pUpper->pExpr->pRight;
dand66e5792016-08-03 16:14:33 +00001867 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n);
1868 if( rc==SQLITE_OK && n ){
danb0b82902014-06-26 20:21:46 +00001869 tRowcnt iNew;
dand66e5792016-08-03 16:14:33 +00001870 u16 mask = WO_GT|WO_LE;
1871 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT);
drh6d3f91d2014-11-05 19:26:12 +00001872 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a);
dand66e5792016-08-03 16:14:33 +00001873 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0);
danb0b82902014-06-26 20:21:46 +00001874 if( iNew<iUpper ) iUpper = iNew;
1875 nOut--;
danf741e042014-08-25 18:29:38 +00001876 pUpper = 0;
danb0b82902014-06-26 20:21:46 +00001877 }
1878 }
1879
1880 pBuilder->pRec = pRec;
1881 if( rc==SQLITE_OK ){
1882 if( iUpper>iLower ){
1883 nNew = sqlite3LogEst(iUpper - iLower);
drh6d3f91d2014-11-05 19:26:12 +00001884 /* TUNING: If both iUpper and iLower are derived from the same
1885 ** sample, then assume they are 4x more selective. This brings
1886 ** the estimated selectivity more in line with what it would be
drh175b8f02019-08-08 15:24:17 +00001887 ** if estimated without the use of STAT4 tables. */
drh6d3f91d2014-11-05 19:26:12 +00001888 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) );
danb0b82902014-06-26 20:21:46 +00001889 }else{
1890 nNew = 10; assert( 10==sqlite3LogEst(2) );
1891 }
1892 if( nNew<nOut ){
1893 nOut = nNew;
1894 }
drhae914d72014-08-28 19:38:22 +00001895 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n",
danb0b82902014-06-26 20:21:46 +00001896 (u32)iLower, (u32)iUpper, nOut));
danb0b82902014-06-26 20:21:46 +00001897 }
1898 }else{
1899 int bDone = 0;
1900 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone);
1901 if( bDone ) return rc;
drh98cdf622009-08-20 18:14:42 +00001902 }
dan02fa4692009-08-17 17:06:58 +00001903 }
drh3f022182009-09-09 16:10:50 +00001904#else
1905 UNUSED_PARAMETER(pParse);
dan7a419232013-08-06 20:01:43 +00001906 UNUSED_PARAMETER(pBuilder);
dan02fa4692009-08-17 17:06:58 +00001907 assert( pLower || pUpper );
danf741e042014-08-25 18:29:38 +00001908#endif
dan7de2a1f2014-04-28 20:11:20 +00001909 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 );
danaa9933c2014-04-24 20:04:49 +00001910 nNew = whereRangeAdjust(pLower, nOut);
1911 nNew = whereRangeAdjust(pUpper, nNew);
dan7de2a1f2014-04-28 20:11:20 +00001912
drh4dd96a82014-10-24 15:26:29 +00001913 /* TUNING: If there is both an upper and lower limit and neither limit
1914 ** has an application-defined likelihood(), assume the range is
dan42685f22014-04-28 19:34:06 +00001915 ** reduced by an additional 75%. This means that, by default, an open-ended
1916 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the
1917 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to
1918 ** match 1/64 of the index. */
drh4dd96a82014-10-24 15:26:29 +00001919 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){
1920 nNew -= 20;
1921 }
dan7de2a1f2014-04-28 20:11:20 +00001922
danaa9933c2014-04-24 20:04:49 +00001923 nOut -= (pLower!=0) + (pUpper!=0);
drhabfa6d52013-09-11 03:53:22 +00001924 if( nNew<10 ) nNew = 10;
1925 if( nNew<nOut ) nOut = nNew;
drhae914d72014-08-28 19:38:22 +00001926#if defined(WHERETRACE_ENABLED)
1927 if( pLoop->nOut>nOut ){
1928 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n",
1929 pLoop->nOut, nOut));
1930 }
1931#endif
drh186ad8c2013-10-08 18:40:37 +00001932 pLoop->nOut = (LogEst)nOut;
dan02fa4692009-08-17 17:06:58 +00001933 return rc;
1934}
1935
drh175b8f02019-08-08 15:24:17 +00001936#ifdef SQLITE_ENABLE_STAT4
drh82759752011-01-20 16:52:09 +00001937/*
1938** Estimate the number of rows that will be returned based on
1939** an equality constraint x=VALUE and where that VALUE occurs in
1940** the histogram data. This only works when x is the left-most
drh175b8f02019-08-08 15:24:17 +00001941** column of an index and sqlite_stat4 histogram data is available
drhac8eb112011-03-17 01:58:21 +00001942** for that index. When pExpr==NULL that means the constraint is
1943** "x IS NULL" instead of "x=VALUE".
drh82759752011-01-20 16:52:09 +00001944**
drh0c50fa02011-01-21 16:27:18 +00001945** Write the estimated row count into *pnRow and return SQLITE_OK.
1946** If unable to make an estimate, leave *pnRow unchanged and return
1947** non-zero.
drh9b3eb0a2011-01-21 14:37:04 +00001948**
1949** This routine can fail if it is unable to load a collating sequence
1950** required for string comparison, or if unable to allocate memory
1951** for a UTF conversion required for comparison. The error is stored
1952** in the pParse structure.
drh82759752011-01-20 16:52:09 +00001953*/
drh041e09f2011-04-07 19:56:21 +00001954static int whereEqualScanEst(
drh82759752011-01-20 16:52:09 +00001955 Parse *pParse, /* Parsing & code generating context */
dan7a419232013-08-06 20:01:43 +00001956 WhereLoopBuilder *pBuilder,
drh0c50fa02011-01-21 16:27:18 +00001957 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */
drhb8a8e8a2013-06-10 19:12:39 +00001958 tRowcnt *pnRow /* Write the revised row estimate here */
drh82759752011-01-20 16:52:09 +00001959){
dan7a419232013-08-06 20:01:43 +00001960 Index *p = pBuilder->pNew->u.btree.pIndex;
1961 int nEq = pBuilder->pNew->u.btree.nEq;
1962 UnpackedRecord *pRec = pBuilder->pRec;
drh82759752011-01-20 16:52:09 +00001963 int rc; /* Subfunction return code */
drhfaacf172011-08-12 01:51:45 +00001964 tRowcnt a[2]; /* Statistics */
dan7a419232013-08-06 20:01:43 +00001965 int bOk;
drh82759752011-01-20 16:52:09 +00001966
dan7a419232013-08-06 20:01:43 +00001967 assert( nEq>=1 );
danfd984b82014-06-30 18:02:20 +00001968 assert( nEq<=p->nColumn );
drh82759752011-01-20 16:52:09 +00001969 assert( p->aSample!=0 );
drh5c624862011-09-22 18:46:34 +00001970 assert( p->nSample>0 );
dan7a419232013-08-06 20:01:43 +00001971 assert( pBuilder->nRecValid<nEq );
1972
1973 /* If values are not available for all fields of the index to the left
1974 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */
1975 if( pBuilder->nRecValid<(nEq-1) ){
1976 return SQLITE_NOTFOUND;
drh1f9c7662011-03-17 01:34:26 +00001977 }
dan7a419232013-08-06 20:01:43 +00001978
dandd6e1f12013-08-10 19:08:30 +00001979 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue()
1980 ** below would return the same value. */
danfd984b82014-06-30 18:02:20 +00001981 if( nEq>=p->nColumn ){
dan7a419232013-08-06 20:01:43 +00001982 *pnRow = 1;
1983 return SQLITE_OK;
drh82759752011-01-20 16:52:09 +00001984 }
dan7a419232013-08-06 20:01:43 +00001985
dand66e5792016-08-03 16:14:33 +00001986 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk);
dan87cd9322013-08-07 15:52:41 +00001987 pBuilder->pRec = pRec;
dan7a419232013-08-06 20:01:43 +00001988 if( rc!=SQLITE_OK ) return rc;
1989 if( bOk==0 ) return SQLITE_NOTFOUND;
dan7a419232013-08-06 20:01:43 +00001990 pBuilder->nRecValid = nEq;
dan7a419232013-08-06 20:01:43 +00001991
danb3c02e22013-08-08 19:38:40 +00001992 whereKeyStats(pParse, p, pRec, 0, a);
drh4fb48e42016-03-01 22:41:27 +00001993 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n",
1994 p->zName, nEq-1, (int)a[1]));
danb3c02e22013-08-08 19:38:40 +00001995 *pnRow = a[1];
daneea568d2013-08-07 19:46:15 +00001996
drh0c50fa02011-01-21 16:27:18 +00001997 return rc;
1998}
drh175b8f02019-08-08 15:24:17 +00001999#endif /* SQLITE_ENABLE_STAT4 */
drh0c50fa02011-01-21 16:27:18 +00002000
drh175b8f02019-08-08 15:24:17 +00002001#ifdef SQLITE_ENABLE_STAT4
drh0c50fa02011-01-21 16:27:18 +00002002/*
2003** Estimate the number of rows that will be returned based on
drh5ac06072011-01-21 18:18:13 +00002004** an IN constraint where the right-hand side of the IN operator
2005** is a list of values. Example:
2006**
2007** WHERE x IN (1,2,3,4)
drh0c50fa02011-01-21 16:27:18 +00002008**
2009** Write the estimated row count into *pnRow and return SQLITE_OK.
2010** If unable to make an estimate, leave *pnRow unchanged and return
2011** non-zero.
2012**
2013** This routine can fail if it is unable to load a collating sequence
2014** required for string comparison, or if unable to allocate memory
2015** for a UTF conversion required for comparison. The error is stored
2016** in the pParse structure.
2017*/
drh041e09f2011-04-07 19:56:21 +00002018static int whereInScanEst(
drh0c50fa02011-01-21 16:27:18 +00002019 Parse *pParse, /* Parsing & code generating context */
dan7a419232013-08-06 20:01:43 +00002020 WhereLoopBuilder *pBuilder,
drh0c50fa02011-01-21 16:27:18 +00002021 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */
drhb8a8e8a2013-06-10 19:12:39 +00002022 tRowcnt *pnRow /* Write the revised row estimate here */
drh0c50fa02011-01-21 16:27:18 +00002023){
dan7a419232013-08-06 20:01:43 +00002024 Index *p = pBuilder->pNew->u.btree.pIndex;
dancfc9df72014-04-25 15:01:01 +00002025 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]);
dan7a419232013-08-06 20:01:43 +00002026 int nRecValid = pBuilder->nRecValid;
drhb8a8e8a2013-06-10 19:12:39 +00002027 int rc = SQLITE_OK; /* Subfunction return code */
2028 tRowcnt nEst; /* Number of rows for a single term */
2029 tRowcnt nRowEst = 0; /* New estimate of the number of rows */
2030 int i; /* Loop counter */
drh0c50fa02011-01-21 16:27:18 +00002031
2032 assert( p->aSample!=0 );
drhfaacf172011-08-12 01:51:45 +00002033 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){
dancfc9df72014-04-25 15:01:01 +00002034 nEst = nRow0;
dan7a419232013-08-06 20:01:43 +00002035 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst);
drhfaacf172011-08-12 01:51:45 +00002036 nRowEst += nEst;
dan7a419232013-08-06 20:01:43 +00002037 pBuilder->nRecValid = nRecValid;
drh0c50fa02011-01-21 16:27:18 +00002038 }
dan7a419232013-08-06 20:01:43 +00002039
drh0c50fa02011-01-21 16:27:18 +00002040 if( rc==SQLITE_OK ){
dancfc9df72014-04-25 15:01:01 +00002041 if( nRowEst > nRow0 ) nRowEst = nRow0;
drh0c50fa02011-01-21 16:27:18 +00002042 *pnRow = nRowEst;
drh5418b122014-08-28 13:42:13 +00002043 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst));
drh0c50fa02011-01-21 16:27:18 +00002044 }
dan7a419232013-08-06 20:01:43 +00002045 assert( pBuilder->nRecValid==nRecValid );
drh0c50fa02011-01-21 16:27:18 +00002046 return rc;
drh82759752011-01-20 16:52:09 +00002047}
drh175b8f02019-08-08 15:24:17 +00002048#endif /* SQLITE_ENABLE_STAT4 */
drh82759752011-01-20 16:52:09 +00002049
drh111a6a72008-12-21 03:51:16 +00002050
drhd15cb172013-05-21 19:23:10 +00002051#ifdef WHERETRACE_ENABLED
drha18f3d22013-05-08 03:05:41 +00002052/*
drhc90713d2014-09-30 13:46:49 +00002053** Print the content of a WhereTerm object
2054*/
drhcacdf202019-12-28 13:39:47 +00002055void sqlite3WhereTermPrint(WhereTerm *pTerm, int iTerm){
drh0a99ba32014-09-30 17:03:35 +00002056 if( pTerm==0 ){
2057 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm);
2058 }else{
drh118efd12019-12-28 14:07:22 +00002059 char zType[8];
drhc84a4022016-05-27 12:30:20 +00002060 char zLeft[50];
drh118efd12019-12-28 14:07:22 +00002061 memcpy(zType, "....", 5);
drh0a99ba32014-09-30 17:03:35 +00002062 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V';
2063 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E';
2064 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L';
drh118efd12019-12-28 14:07:22 +00002065 if( pTerm->wtFlags & TERM_CODED ) zType[3] = 'C';
drhc84a4022016-05-27 12:30:20 +00002066 if( pTerm->eOperator & WO_SINGLE ){
drh220f0d62021-10-15 17:06:16 +00002067 assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 );
drhc84a4022016-05-27 12:30:20 +00002068 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}",
drh75fa2662020-09-28 15:49:43 +00002069 pTerm->leftCursor, pTerm->u.x.leftColumn);
drhc84a4022016-05-27 12:30:20 +00002070 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){
drhe93986a2020-10-17 19:09:04 +00002071 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%llx",
drhc84a4022016-05-27 12:30:20 +00002072 pTerm->u.pOrInfo->indexable);
2073 }else{
2074 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor);
2075 }
drhfcd49532015-05-13 15:24:07 +00002076 sqlite3DebugPrintf(
drh118efd12019-12-28 14:07:22 +00002077 "TERM-%-3d %p %s %-12s op=%03x wtFlags=%04x",
2078 iTerm, pTerm, zType, zLeft, pTerm->eOperator, pTerm->wtFlags);
2079 /* The 0x10000 .wheretrace flag causes extra information to be
2080 ** shown about each Term */
drh6411d652019-12-28 12:33:35 +00002081 if( sqlite3WhereTrace & 0x10000 ){
drh118efd12019-12-28 14:07:22 +00002082 sqlite3DebugPrintf(" prob=%-3d prereq=%llx,%llx",
2083 pTerm->truthProb, (u64)pTerm->prereqAll, (u64)pTerm->prereqRight);
drh6411d652019-12-28 12:33:35 +00002084 }
drh220f0d62021-10-15 17:06:16 +00002085 if( (pTerm->eOperator & (WO_OR|WO_AND))==0 && pTerm->u.x.iField ){
drh75fa2662020-09-28 15:49:43 +00002086 sqlite3DebugPrintf(" iField=%d", pTerm->u.x.iField);
drha15a8bc2016-08-19 18:40:17 +00002087 }
drhd262c2d2019-12-22 19:41:12 +00002088 if( pTerm->iParent>=0 ){
2089 sqlite3DebugPrintf(" iParent=%d", pTerm->iParent);
2090 }
2091 sqlite3DebugPrintf("\n");
drh0a99ba32014-09-30 17:03:35 +00002092 sqlite3TreeViewExpr(0, pTerm->pExpr, 0);
2093 }
drhc90713d2014-09-30 13:46:49 +00002094}
2095#endif
2096
2097#ifdef WHERETRACE_ENABLED
2098/*
drhc84a4022016-05-27 12:30:20 +00002099** Show the complete content of a WhereClause
2100*/
2101void sqlite3WhereClausePrint(WhereClause *pWC){
2102 int i;
2103 for(i=0; i<pWC->nTerm; i++){
drhcacdf202019-12-28 13:39:47 +00002104 sqlite3WhereTermPrint(&pWC->a[i], i);
drhc84a4022016-05-27 12:30:20 +00002105 }
2106}
2107#endif
2108
2109#ifdef WHERETRACE_ENABLED
2110/*
drha18f3d22013-05-08 03:05:41 +00002111** Print a WhereLoop object for debugging purposes
2112*/
drhcacdf202019-12-28 13:39:47 +00002113void sqlite3WhereLoopPrint(WhereLoop *p, WhereClause *pWC){
drhc1ba2e72013-10-28 19:03:21 +00002114 WhereInfo *pWInfo = pWC->pWInfo;
drh53801ef2016-04-09 14:36:07 +00002115 int nb = 1+(pWInfo->pTabList->nSrc+3)/4;
drh76012942021-02-21 21:04:54 +00002116 SrcItem *pItem = pWInfo->pTabList->a + p->iTab;
drha18f3d22013-05-08 03:05:41 +00002117 Table *pTab = pItem->pTab;
drh53801ef2016-04-09 14:36:07 +00002118 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1;
drh6457a352013-06-21 00:35:37 +00002119 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId,
drh53801ef2016-04-09 14:36:07 +00002120 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll);
drh6457a352013-06-21 00:35:37 +00002121 sqlite3DebugPrintf(" %12s",
drha18f3d22013-05-08 03:05:41 +00002122 pItem->zAlias ? pItem->zAlias : pTab->zName);
drh5346e952013-05-08 14:14:26 +00002123 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){
drhf3f69ac2014-08-20 23:38:07 +00002124 const char *zName;
2125 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){
drh319f6772013-05-14 15:31:07 +00002126 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){
2127 int i = sqlite3Strlen30(zName) - 1;
2128 while( zName[i]!='_' ) i--;
2129 zName += i;
2130 }
drh6457a352013-06-21 00:35:37 +00002131 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq);
drh5346e952013-05-08 14:14:26 +00002132 }else{
drh6457a352013-06-21 00:35:37 +00002133 sqlite3DebugPrintf("%20s","");
drh5346e952013-05-08 14:14:26 +00002134 }
drha18f3d22013-05-08 03:05:41 +00002135 }else{
drh5346e952013-05-08 14:14:26 +00002136 char *z;
2137 if( p->u.vtab.idxStr ){
drh05fbfd82019-12-05 17:31:58 +00002138 z = sqlite3_mprintf("(%d,\"%s\",%#x)",
drh3bd26f02013-05-24 14:52:03 +00002139 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask);
drh5346e952013-05-08 14:14:26 +00002140 }else{
drh3bd26f02013-05-24 14:52:03 +00002141 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask);
drh5346e952013-05-08 14:14:26 +00002142 }
drh6457a352013-06-21 00:35:37 +00002143 sqlite3DebugPrintf(" %-19s", z);
drh5346e952013-05-08 14:14:26 +00002144 sqlite3_free(z);
drha18f3d22013-05-08 03:05:41 +00002145 }
drhf3f69ac2014-08-20 23:38:07 +00002146 if( p->wsFlags & WHERE_SKIPSCAN ){
drhfb82caf2021-12-08 19:50:45 +00002147 sqlite3DebugPrintf(" f %06x %d-%d", p->wsFlags, p->nLTerm,p->nSkip);
drhf3f69ac2014-08-20 23:38:07 +00002148 }else{
drhfb82caf2021-12-08 19:50:45 +00002149 sqlite3DebugPrintf(" f %06x N %d", p->wsFlags, p->nLTerm);
drhf3f69ac2014-08-20 23:38:07 +00002150 }
drhb8a8e8a2013-06-10 19:12:39 +00002151 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut);
drhc90713d2014-09-30 13:46:49 +00002152 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){
2153 int i;
2154 for(i=0; i<p->nLTerm; i++){
drhcacdf202019-12-28 13:39:47 +00002155 sqlite3WhereTermPrint(p->aLTerm[i], i);
drhc90713d2014-09-30 13:46:49 +00002156 }
2157 }
drha18f3d22013-05-08 03:05:41 +00002158}
2159#endif
2160
drhf1b5f5b2013-05-02 00:15:01 +00002161/*
drh4efc9292013-06-06 23:02:03 +00002162** Convert bulk memory into a valid WhereLoop that can be passed
2163** to whereLoopClear harmlessly.
drh5346e952013-05-08 14:14:26 +00002164*/
drh4efc9292013-06-06 23:02:03 +00002165static void whereLoopInit(WhereLoop *p){
2166 p->aLTerm = p->aLTermSpace;
2167 p->nLTerm = 0;
2168 p->nLSlot = ArraySize(p->aLTermSpace);
2169 p->wsFlags = 0;
2170}
2171
2172/*
2173** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact.
2174*/
2175static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){
drh986b3872013-06-28 21:12:20 +00002176 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){
drh13e11b42013-06-06 23:44:25 +00002177 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){
2178 sqlite3_free(p->u.vtab.idxStr);
2179 p->u.vtab.needFree = 0;
2180 p->u.vtab.idxStr = 0;
drh986b3872013-06-28 21:12:20 +00002181 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){
drh13e11b42013-06-06 23:44:25 +00002182 sqlite3DbFree(db, p->u.btree.pIndex->zColAff);
drhdbd6a7d2017-04-05 12:39:49 +00002183 sqlite3DbFreeNN(db, p->u.btree.pIndex);
drh13e11b42013-06-06 23:44:25 +00002184 p->u.btree.pIndex = 0;
2185 }
drh5346e952013-05-08 14:14:26 +00002186 }
2187}
2188
drh4efc9292013-06-06 23:02:03 +00002189/*
2190** Deallocate internal memory used by a WhereLoop object
2191*/
2192static void whereLoopClear(sqlite3 *db, WhereLoop *p){
drhdbd6a7d2017-04-05 12:39:49 +00002193 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm);
drh4efc9292013-06-06 23:02:03 +00002194 whereLoopClearUnion(db, p);
2195 whereLoopInit(p);
2196}
2197
2198/*
2199** Increase the memory allocation for pLoop->aLTerm[] to be at least n.
2200*/
2201static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){
2202 WhereTerm **paNew;
2203 if( p->nLSlot>=n ) return SQLITE_OK;
2204 n = (n+7)&~7;
drh575fad62016-02-05 13:38:36 +00002205 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n);
mistachkinfad30392016-02-13 23:43:46 +00002206 if( paNew==0 ) return SQLITE_NOMEM_BKPT;
drh4efc9292013-06-06 23:02:03 +00002207 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot);
drhdbd6a7d2017-04-05 12:39:49 +00002208 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm);
drh4efc9292013-06-06 23:02:03 +00002209 p->aLTerm = paNew;
2210 p->nLSlot = n;
2211 return SQLITE_OK;
2212}
2213
2214/*
2215** Transfer content from the second pLoop into the first.
2216*/
2217static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){
drh4efc9292013-06-06 23:02:03 +00002218 whereLoopClearUnion(db, pTo);
drh0d31dc32013-09-06 00:40:59 +00002219 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){
drh82404312021-04-22 12:38:30 +00002220 memset(pTo, 0, WHERE_LOOP_XFER_SZ);
mistachkinfad30392016-02-13 23:43:46 +00002221 return SQLITE_NOMEM_BKPT;
drh0d31dc32013-09-06 00:40:59 +00002222 }
drha2014152013-06-07 00:29:23 +00002223 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ);
2224 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0]));
drh4efc9292013-06-06 23:02:03 +00002225 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){
2226 pFrom->u.vtab.needFree = 0;
drh986b3872013-06-28 21:12:20 +00002227 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){
drh4efc9292013-06-06 23:02:03 +00002228 pFrom->u.btree.pIndex = 0;
2229 }
2230 return SQLITE_OK;
2231}
2232
drh5346e952013-05-08 14:14:26 +00002233/*
drhf1b5f5b2013-05-02 00:15:01 +00002234** Delete a WhereLoop object
2235*/
2236static void whereLoopDelete(sqlite3 *db, WhereLoop *p){
drh5346e952013-05-08 14:14:26 +00002237 whereLoopClear(db, p);
drhdbd6a7d2017-04-05 12:39:49 +00002238 sqlite3DbFreeNN(db, p);
drhf1b5f5b2013-05-02 00:15:01 +00002239}
drh84bfda42005-07-15 13:05:21 +00002240
drh9eff6162006-06-12 21:59:13 +00002241/*
2242** Free a WhereInfo structure
2243*/
drh10fe8402008-10-11 16:47:35 +00002244static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){
drh9d9c41e2017-10-31 03:40:15 +00002245 assert( pWInfo!=0 );
drh9d9c41e2017-10-31 03:40:15 +00002246 sqlite3WhereClauseClear(&pWInfo->sWC);
2247 while( pWInfo->pLoops ){
2248 WhereLoop *p = pWInfo->pLoops;
2249 pWInfo->pLoops = p->pNextLoop;
2250 whereLoopDelete(db, p);
2251 }
drh36e678b2020-01-02 00:45:38 +00002252 assert( pWInfo->pExprMods==0 );
drhf8bdcfa2022-04-09 03:06:01 +00002253 while( pWInfo->pMemToFree ){
2254 WhereMemBlock *pNext = pWInfo->pMemToFree->pNext;
2255 sqlite3DbFreeNN(db, pWInfo->pMemToFree);
2256 pWInfo->pMemToFree = pNext;
2257 }
drh9d9c41e2017-10-31 03:40:15 +00002258 sqlite3DbFreeNN(db, pWInfo);
drh9eff6162006-06-12 21:59:13 +00002259}
2260
drhd784cc82021-04-15 12:56:44 +00002261/* Undo all Expr node modifications
2262*/
2263static void whereUndoExprMods(WhereInfo *pWInfo){
2264 while( pWInfo->pExprMods ){
2265 WhereExprMod *p = pWInfo->pExprMods;
2266 pWInfo->pExprMods = p->pNext;
2267 memcpy(p->pExpr, &p->orig, sizeof(p->orig));
2268 sqlite3DbFree(pWInfo->pParse->db, p);
2269 }
2270}
2271
drhf1b5f5b2013-05-02 00:15:01 +00002272/*
drhe0de8762014-11-05 13:13:13 +00002273** Return TRUE if all of the following are true:
drhb355c2c2014-04-18 22:20:31 +00002274**
dan748d8b92021-08-31 15:53:58 +00002275** (1) X has the same or lower cost, or returns the same or fewer rows,
2276** than Y.
drh989d7272017-10-16 11:50:12 +00002277** (2) X uses fewer WHERE clause terms than Y
drh47b1d682017-10-15 22:16:25 +00002278** (3) Every WHERE clause term used by X is also used by Y
2279** (4) X skips at least as many columns as Y
2280** (5) If X is a covering index, than Y is too
drhb355c2c2014-04-18 22:20:31 +00002281**
drh47b1d682017-10-15 22:16:25 +00002282** Conditions (2) and (3) mean that X is a "proper subset" of Y.
drhb355c2c2014-04-18 22:20:31 +00002283** If X is a proper subset of Y then Y is a better choice and ought
2284** to have a lower cost. This routine returns TRUE when that cost
drh47b1d682017-10-15 22:16:25 +00002285** relationship is inverted and needs to be adjusted. Constraint (4)
drhe0de8762014-11-05 13:13:13 +00002286** was added because if X uses skip-scan less than Y it still might
drh989d7272017-10-16 11:50:12 +00002287** deserve a lower cost even if it is a proper subset of Y. Constraint (5)
2288** was added because a covering index probably deserves to have a lower cost
2289** than a non-covering index even if it is a proper subset.
drh3fb183d2014-03-31 19:49:00 +00002290*/
drhb355c2c2014-04-18 22:20:31 +00002291static int whereLoopCheaperProperSubset(
2292 const WhereLoop *pX, /* First WhereLoop to compare */
2293 const WhereLoop *pY /* Compare against this WhereLoop */
2294){
drh3fb183d2014-03-31 19:49:00 +00002295 int i, j;
drhc8bbce12014-10-21 01:05:09 +00002296 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){
2297 return 0; /* X is not a subset of Y */
2298 }
dan748d8b92021-08-31 15:53:58 +00002299 if( pX->rRun>pY->rRun && pX->nOut>pY->nOut ) return 0;
drhe0de8762014-11-05 13:13:13 +00002300 if( pY->nSkip > pX->nSkip ) return 0;
drh9ee88102014-05-07 20:33:17 +00002301 for(i=pX->nLTerm-1; i>=0; i--){
drhc8bbce12014-10-21 01:05:09 +00002302 if( pX->aLTerm[i]==0 ) continue;
drhb355c2c2014-04-18 22:20:31 +00002303 for(j=pY->nLTerm-1; j>=0; j--){
2304 if( pY->aLTerm[j]==pX->aLTerm[i] ) break;
2305 }
2306 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */
2307 }
drh47b1d682017-10-15 22:16:25 +00002308 if( (pX->wsFlags&WHERE_IDX_ONLY)!=0
2309 && (pY->wsFlags&WHERE_IDX_ONLY)==0 ){
2310 return 0; /* Constraint (5) */
2311 }
drhb355c2c2014-04-18 22:20:31 +00002312 return 1; /* All conditions meet */
drh3fb183d2014-03-31 19:49:00 +00002313}
2314
2315/*
dan748d8b92021-08-31 15:53:58 +00002316** Try to adjust the cost and number of output rows of WhereLoop pTemplate
2317** upwards or downwards so that:
drh53cd10a2014-03-31 18:24:18 +00002318**
drh3fb183d2014-03-31 19:49:00 +00002319** (1) pTemplate costs less than any other WhereLoops that are a proper
2320** subset of pTemplate
drh53cd10a2014-03-31 18:24:18 +00002321**
drh3fb183d2014-03-31 19:49:00 +00002322** (2) pTemplate costs more than any other WhereLoops for which pTemplate
2323** is a proper subset.
drh53cd10a2014-03-31 18:24:18 +00002324**
drh3fb183d2014-03-31 19:49:00 +00002325** To say "WhereLoop X is a proper subset of Y" means that X uses fewer
2326** WHERE clause terms than Y and that every WHERE clause term used by X is
2327** also used by Y.
drh53cd10a2014-03-31 18:24:18 +00002328*/
2329static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){
2330 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return;
drh53cd10a2014-03-31 18:24:18 +00002331 for(; p; p=p->pNextLoop){
drh3fb183d2014-03-31 19:49:00 +00002332 if( p->iTab!=pTemplate->iTab ) continue;
2333 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue;
drhb355c2c2014-04-18 22:20:31 +00002334 if( whereLoopCheaperProperSubset(p, pTemplate) ){
2335 /* Adjust pTemplate cost downward so that it is cheaper than its
drhe0de8762014-11-05 13:13:13 +00002336 ** subset p. */
drh1b131b72014-10-21 16:01:40 +00002337 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n",
dan748d8b92021-08-31 15:53:58 +00002338 pTemplate->rRun, pTemplate->nOut,
2339 MIN(p->rRun, pTemplate->rRun),
2340 MIN(p->nOut - 1, pTemplate->nOut)));
2341 pTemplate->rRun = MIN(p->rRun, pTemplate->rRun);
2342 pTemplate->nOut = MIN(p->nOut - 1, pTemplate->nOut);
drhb355c2c2014-04-18 22:20:31 +00002343 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){
2344 /* Adjust pTemplate cost upward so that it is costlier than p since
2345 ** pTemplate is a proper subset of p */
drh1b131b72014-10-21 16:01:40 +00002346 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n",
dan748d8b92021-08-31 15:53:58 +00002347 pTemplate->rRun, pTemplate->nOut,
2348 MAX(p->rRun, pTemplate->rRun),
2349 MAX(p->nOut + 1, pTemplate->nOut)));
2350 pTemplate->rRun = MAX(p->rRun, pTemplate->rRun);
2351 pTemplate->nOut = MAX(p->nOut + 1, pTemplate->nOut);
drh53cd10a2014-03-31 18:24:18 +00002352 }
2353 }
2354}
2355
2356/*
drh7a4b1642014-03-29 21:16:07 +00002357** Search the list of WhereLoops in *ppPrev looking for one that can be
drhbcbb0662017-05-19 20:55:04 +00002358** replaced by pTemplate.
drhf1b5f5b2013-05-02 00:15:01 +00002359**
drhbcbb0662017-05-19 20:55:04 +00002360** Return NULL if pTemplate does not belong on the WhereLoop list.
2361** In other words if pTemplate ought to be dropped from further consideration.
drh23f98da2013-05-21 15:52:07 +00002362**
drhbcbb0662017-05-19 20:55:04 +00002363** If pX is a WhereLoop that pTemplate can replace, then return the
drh7a4b1642014-03-29 21:16:07 +00002364** link that points to pX.
drh23f98da2013-05-21 15:52:07 +00002365**
drhbcbb0662017-05-19 20:55:04 +00002366** If pTemplate cannot replace any existing element of the list but needs
2367** to be added to the list as a new entry, then return a pointer to the
2368** tail of the list.
drhf1b5f5b2013-05-02 00:15:01 +00002369*/
drh7a4b1642014-03-29 21:16:07 +00002370static WhereLoop **whereLoopFindLesser(
2371 WhereLoop **ppPrev,
2372 const WhereLoop *pTemplate
2373){
2374 WhereLoop *p;
2375 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){
drhdbb80232013-06-19 12:34:13 +00002376 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){
2377 /* If either the iTab or iSortIdx values for two WhereLoop are different
2378 ** then those WhereLoops need to be considered separately. Neither is
2379 ** a candidate to replace the other. */
2380 continue;
2381 }
2382 /* In the current implementation, the rSetup value is either zero
2383 ** or the cost of building an automatic index (NlogN) and the NlogN
2384 ** is the same for compatible WhereLoops. */
2385 assert( p->rSetup==0 || pTemplate->rSetup==0
2386 || p->rSetup==pTemplate->rSetup );
2387
2388 /* whereLoopAddBtree() always generates and inserts the automatic index
2389 ** case first. Hence compatible candidate WhereLoops never have a larger
2390 ** rSetup. Call this SETUP-INVARIANT */
2391 assert( p->rSetup>=pTemplate->rSetup );
2392
drhdabe36d2014-06-17 20:16:43 +00002393 /* Any loop using an appliation-defined index (or PRIMARY KEY or
2394 ** UNIQUE constraint) with one or more == constraints is better
dan70273d02014-11-14 19:34:20 +00002395 ** than an automatic index. Unless it is a skip-scan. */
drhdabe36d2014-06-17 20:16:43 +00002396 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0
dan70273d02014-11-14 19:34:20 +00002397 && (pTemplate->nSkip)==0
drhdabe36d2014-06-17 20:16:43 +00002398 && (pTemplate->wsFlags & WHERE_INDEXED)!=0
2399 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0
2400 && (p->prereq & pTemplate->prereq)==pTemplate->prereq
2401 ){
2402 break;
2403 }
2404
drh53cd10a2014-03-31 18:24:18 +00002405 /* If existing WhereLoop p is better than pTemplate, pTemplate can be
2406 ** discarded. WhereLoop p is better if:
2407 ** (1) p has no more dependencies than pTemplate, and
2408 ** (2) p has an equal or lower cost than pTemplate
2409 */
2410 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */
2411 && p->rSetup<=pTemplate->rSetup /* (2a) */
2412 && p->rRun<=pTemplate->rRun /* (2b) */
2413 && p->nOut<=pTemplate->nOut /* (2c) */
drhf1b5f5b2013-05-02 00:15:01 +00002414 ){
drh53cd10a2014-03-31 18:24:18 +00002415 return 0; /* Discard pTemplate */
drhf1b5f5b2013-05-02 00:15:01 +00002416 }
drh53cd10a2014-03-31 18:24:18 +00002417
2418 /* If pTemplate is always better than p, then cause p to be overwritten
2419 ** with pTemplate. pTemplate is better than p if:
2420 ** (1) pTemplate has no more dependences than p, and
2421 ** (2) pTemplate has an equal or lower cost than p.
2422 */
2423 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */
2424 && p->rRun>=pTemplate->rRun /* (2a) */
2425 && p->nOut>=pTemplate->nOut /* (2b) */
drhf1b5f5b2013-05-02 00:15:01 +00002426 ){
drhadd5ce32013-09-07 00:29:06 +00002427 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */
drh53cd10a2014-03-31 18:24:18 +00002428 break; /* Cause p to be overwritten by pTemplate */
drhf1b5f5b2013-05-02 00:15:01 +00002429 }
2430 }
drh7a4b1642014-03-29 21:16:07 +00002431 return ppPrev;
2432}
2433
2434/*
drh94a11212004-09-25 13:12:14 +00002435** Insert or replace a WhereLoop entry using the template supplied.
2436**
2437** An existing WhereLoop entry might be overwritten if the new template
2438** is better and has fewer dependencies. Or the template will be ignored
2439** and no insert will occur if an existing WhereLoop is faster and has
2440** fewer dependencies than the template. Otherwise a new WhereLoop is
danielk1977b3bce662005-01-29 08:32:43 +00002441** added based on the template.
drh94a11212004-09-25 13:12:14 +00002442**
drh7a4b1642014-03-29 21:16:07 +00002443** If pBuilder->pOrSet is not NULL then we care about only the
drh94a11212004-09-25 13:12:14 +00002444** prerequisites and rRun and nOut costs of the N best loops. That
2445** information is gathered in the pBuilder->pOrSet object. This special
2446** processing mode is used only for OR clause processing.
danielk1977b3bce662005-01-29 08:32:43 +00002447**
drh94a11212004-09-25 13:12:14 +00002448** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we
danielk1977b3bce662005-01-29 08:32:43 +00002449** still might overwrite similar loops with the new template if the
drh53cd10a2014-03-31 18:24:18 +00002450** new template is better. Loops may be overwritten if the following
drh94a11212004-09-25 13:12:14 +00002451** conditions are met:
2452**
2453** (1) They have the same iTab.
2454** (2) They have the same iSortIdx.
2455** (3) The template has same or fewer dependencies than the current loop
2456** (4) The template has the same or lower cost than the current loop
drh94a11212004-09-25 13:12:14 +00002457*/
2458static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){
drh7a4b1642014-03-29 21:16:07 +00002459 WhereLoop **ppPrev, *p;
drh94a11212004-09-25 13:12:14 +00002460 WhereInfo *pWInfo = pBuilder->pWInfo;
2461 sqlite3 *db = pWInfo->pParse->db;
drhbacbbcc2016-03-09 12:35:18 +00002462 int rc;
drh94a11212004-09-25 13:12:14 +00002463
drhfc9098a2018-09-21 18:43:51 +00002464 /* Stop the search once we hit the query planner search limit */
drh2c3ba942018-09-22 15:05:32 +00002465 if( pBuilder->iPlanLimit==0 ){
2466 WHERETRACE(0xffffffff,("=== query planner search limit reached ===\n"));
2467 if( pBuilder->pOrSet ) pBuilder->pOrSet->n = 0;
2468 return SQLITE_DONE;
2469 }
drhfc9098a2018-09-21 18:43:51 +00002470 pBuilder->iPlanLimit--;
2471
dan51f2b172019-12-28 15:24:02 +00002472 whereLoopAdjustCost(pWInfo->pLoops, pTemplate);
2473
drh94a11212004-09-25 13:12:14 +00002474 /* If pBuilder->pOrSet is defined, then only keep track of the costs
2475 ** and prereqs.
2476 */
2477 if( pBuilder->pOrSet!=0 ){
drh2dc29292015-08-27 23:18:55 +00002478 if( pTemplate->nLTerm ){
drh94a11212004-09-25 13:12:14 +00002479#if WHERETRACE_ENABLED
drh2dc29292015-08-27 23:18:55 +00002480 u16 n = pBuilder->pOrSet->n;
2481 int x =
drh94a11212004-09-25 13:12:14 +00002482#endif
drh2dc29292015-08-27 23:18:55 +00002483 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun,
drh94a11212004-09-25 13:12:14 +00002484 pTemplate->nOut);
2485#if WHERETRACE_ENABLED /* 0x8 */
drh2dc29292015-08-27 23:18:55 +00002486 if( sqlite3WhereTrace & 0x8 ){
2487 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n);
drhcacdf202019-12-28 13:39:47 +00002488 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC);
drh2dc29292015-08-27 23:18:55 +00002489 }
drh94a11212004-09-25 13:12:14 +00002490#endif
drh2dc29292015-08-27 23:18:55 +00002491 }
danielk1977b3bce662005-01-29 08:32:43 +00002492 return SQLITE_OK;
drh9012bcb2004-12-19 00:11:35 +00002493 }
drh94a11212004-09-25 13:12:14 +00002494
drh7a4b1642014-03-29 21:16:07 +00002495 /* Look for an existing WhereLoop to replace with pTemplate
drh51669862004-12-18 18:40:26 +00002496 */
drh7a4b1642014-03-29 21:16:07 +00002497 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate);
drhf1b5f5b2013-05-02 00:15:01 +00002498
drh7a4b1642014-03-29 21:16:07 +00002499 if( ppPrev==0 ){
2500 /* There already exists a WhereLoop on the list that is better
2501 ** than pTemplate, so just ignore pTemplate */
2502#if WHERETRACE_ENABLED /* 0x8 */
2503 if( sqlite3WhereTrace & 0x8 ){
drh9a7b41d2014-10-08 00:08:08 +00002504 sqlite3DebugPrintf(" skip: ");
drhcacdf202019-12-28 13:39:47 +00002505 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC);
drhf1b5f5b2013-05-02 00:15:01 +00002506 }
drh7a4b1642014-03-29 21:16:07 +00002507#endif
2508 return SQLITE_OK;
2509 }else{
2510 p = *ppPrev;
drhf1b5f5b2013-05-02 00:15:01 +00002511 }
2512
2513 /* If we reach this point it means that either p[] should be overwritten
2514 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new
2515 ** WhereLoop and insert it.
2516 */
drh989578e2013-10-28 14:34:35 +00002517#if WHERETRACE_ENABLED /* 0x8 */
drhae70cf12013-05-31 15:18:46 +00002518 if( sqlite3WhereTrace & 0x8 ){
2519 if( p!=0 ){
drh9a7b41d2014-10-08 00:08:08 +00002520 sqlite3DebugPrintf("replace: ");
drhcacdf202019-12-28 13:39:47 +00002521 sqlite3WhereLoopPrint(p, pBuilder->pWC);
drhbcbb0662017-05-19 20:55:04 +00002522 sqlite3DebugPrintf(" with: ");
2523 }else{
2524 sqlite3DebugPrintf(" add: ");
drhae70cf12013-05-31 15:18:46 +00002525 }
drhcacdf202019-12-28 13:39:47 +00002526 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC);
drhae70cf12013-05-31 15:18:46 +00002527 }
2528#endif
drhf1b5f5b2013-05-02 00:15:01 +00002529 if( p==0 ){
drh7a4b1642014-03-29 21:16:07 +00002530 /* Allocate a new WhereLoop to add to the end of the list */
drh575fad62016-02-05 13:38:36 +00002531 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop));
mistachkinfad30392016-02-13 23:43:46 +00002532 if( p==0 ) return SQLITE_NOMEM_BKPT;
drh4efc9292013-06-06 23:02:03 +00002533 whereLoopInit(p);
drh7a4b1642014-03-29 21:16:07 +00002534 p->pNextLoop = 0;
2535 }else{
2536 /* We will be overwriting WhereLoop p[]. But before we do, first
2537 ** go through the rest of the list and delete any other entries besides
2538 ** p[] that are also supplated by pTemplate */
2539 WhereLoop **ppTail = &p->pNextLoop;
2540 WhereLoop *pToDel;
2541 while( *ppTail ){
2542 ppTail = whereLoopFindLesser(ppTail, pTemplate);
drhdabe36d2014-06-17 20:16:43 +00002543 if( ppTail==0 ) break;
drh7a4b1642014-03-29 21:16:07 +00002544 pToDel = *ppTail;
2545 if( pToDel==0 ) break;
2546 *ppTail = pToDel->pNextLoop;
2547#if WHERETRACE_ENABLED /* 0x8 */
2548 if( sqlite3WhereTrace & 0x8 ){
drh9a7b41d2014-10-08 00:08:08 +00002549 sqlite3DebugPrintf(" delete: ");
drhcacdf202019-12-28 13:39:47 +00002550 sqlite3WhereLoopPrint(pToDel, pBuilder->pWC);
drh7a4b1642014-03-29 21:16:07 +00002551 }
2552#endif
2553 whereLoopDelete(db, pToDel);
2554 }
drhf1b5f5b2013-05-02 00:15:01 +00002555 }
drhbacbbcc2016-03-09 12:35:18 +00002556 rc = whereLoopXfer(db, p, pTemplate);
drh5346e952013-05-08 14:14:26 +00002557 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){
drhef866372013-05-22 20:49:02 +00002558 Index *pIndex = p->u.btree.pIndex;
drh5f913ec2019-01-10 13:56:08 +00002559 if( pIndex && pIndex->idxType==SQLITE_IDXTYPE_IPK ){
drhcf8fa7a2013-05-10 20:26:22 +00002560 p->u.btree.pIndex = 0;
2561 }
drh5346e952013-05-08 14:14:26 +00002562 }
drhbacbbcc2016-03-09 12:35:18 +00002563 return rc;
drhf1b5f5b2013-05-02 00:15:01 +00002564}
2565
2566/*
drhcca9f3d2013-09-06 15:23:29 +00002567** Adjust the WhereLoop.nOut value downward to account for terms of the
2568** WHERE clause that reference the loop but which are not used by an
2569** index.
drh7a1bca72014-11-22 18:50:44 +00002570*
2571** For every WHERE clause term that is not used by the index
2572** and which has a truth probability assigned by one of the likelihood(),
2573** likely(), or unlikely() SQL functions, reduce the estimated number
2574** of output rows by the probability specified.
drhcca9f3d2013-09-06 15:23:29 +00002575**
drh7a1bca72014-11-22 18:50:44 +00002576** TUNING: For every WHERE clause term that is not used by the index
2577** and which does not have an assigned truth probability, heuristics
2578** described below are used to try to estimate the truth probability.
2579** TODO --> Perhaps this is something that could be improved by better
2580** table statistics.
2581**
drhab4624d2014-11-22 19:52:10 +00002582** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75%
2583** value corresponds to -1 in LogEst notation, so this means decrement
drh7a1bca72014-11-22 18:50:44 +00002584** the WhereLoop.nOut field for every such WHERE clause term.
2585**
2586** Heuristic 2: If there exists one or more WHERE clause terms of the
2587** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the
2588** final output row estimate is no greater than 1/4 of the total number
2589** of rows in the table. In other words, assume that x==EXPR will filter
2590** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the
2591** "x" column is boolean or else -1 or 0 or 1 is a common default value
2592** on the "x" column and so in that case only cap the output row estimate
2593** at 1/2 instead of 1/4.
drhcca9f3d2013-09-06 15:23:29 +00002594*/
drhd8b77e22014-09-06 01:35:57 +00002595static void whereLoopOutputAdjust(
2596 WhereClause *pWC, /* The WHERE clause */
2597 WhereLoop *pLoop, /* The loop to adjust downward */
2598 LogEst nRow /* Number of rows in the entire table */
2599){
drh7d9e7d82013-09-11 17:39:09 +00002600 WhereTerm *pTerm, *pX;
drhcca9f3d2013-09-06 15:23:29 +00002601 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf);
drhaa16c602019-09-18 12:49:34 +00002602 int i, j;
drh7a1bca72014-11-22 18:50:44 +00002603 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */
drhadd5ce32013-09-07 00:29:06 +00002604
drha3898252014-11-22 12:22:13 +00002605 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 );
drh132f96f2021-12-08 16:07:22 +00002606 for(i=pWC->nBase, pTerm=pWC->a; i>0; i--, pTerm++){
drh55f66b32019-07-16 19:44:32 +00002607 assert( pTerm!=0 );
drhcca9f3d2013-09-06 15:23:29 +00002608 if( (pTerm->prereqAll & notAllowed)!=0 ) continue;
drh132f96f2021-12-08 16:07:22 +00002609 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue;
2610 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) continue;
drh7d9e7d82013-09-11 17:39:09 +00002611 for(j=pLoop->nLTerm-1; j>=0; j--){
2612 pX = pLoop->aLTerm[j];
drhd2447442013-11-13 19:01:41 +00002613 if( pX==0 ) continue;
drh7d9e7d82013-09-11 17:39:09 +00002614 if( pX==pTerm ) break;
2615 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break;
2616 }
danaa9933c2014-04-24 20:04:49 +00002617 if( j<0 ){
drhfb82caf2021-12-08 19:50:45 +00002618 if( pLoop->maskSelf==pTerm->prereqAll ){
drh7e910f62021-12-09 01:28:15 +00002619 /* If there are extra terms in the WHERE clause not used by an index
2620 ** that depend only on the table being scanned, and that will tend to
2621 ** cause many rows to be omitted, then mark that table as
drhbe341502022-03-25 01:23:37 +00002622 ** "self-culling".
2623 **
2624 ** 2022-03-24: Self-culling only applies if either the extra terms
2625 ** are straight comparison operators that are non-true with NULL
drha76ac882022-04-08 19:20:12 +00002626 ** operand, or if the loop is not an OUTER JOIN.
drhbe341502022-03-25 01:23:37 +00002627 */
2628 if( (pTerm->eOperator & 0x3f)!=0
drha76ac882022-04-08 19:20:12 +00002629 || (pWC->pWInfo->pTabList->a[pLoop->iTab].fg.jointype
2630 & (JT_LEFT|JT_LTORJ))==0
drhbe341502022-03-25 01:23:37 +00002631 ){
2632 pLoop->wsFlags |= WHERE_SELFCULL;
2633 }
drhfb82caf2021-12-08 19:50:45 +00002634 }
drhd8b77e22014-09-06 01:35:57 +00002635 if( pTerm->truthProb<=0 ){
drh7a1bca72014-11-22 18:50:44 +00002636 /* If a truth probability is specified using the likelihood() hints,
2637 ** then use the probability provided by the application. */
drhd8b77e22014-09-06 01:35:57 +00002638 pLoop->nOut += pTerm->truthProb;
2639 }else{
drh7a1bca72014-11-22 18:50:44 +00002640 /* In the absence of explicit truth probabilities, use heuristics to
2641 ** guess a reasonable truth probability. */
drhd8b77e22014-09-06 01:35:57 +00002642 pLoop->nOut--;
drhf06cdde2020-02-24 16:46:08 +00002643 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0
2644 && (pTerm->wtFlags & TERM_HIGHTRUTH)==0 /* tag-20200224-1 */
2645 ){
drh7a1bca72014-11-22 18:50:44 +00002646 Expr *pRight = pTerm->pExpr->pRight;
drhaa16c602019-09-18 12:49:34 +00002647 int k = 0;
drhe0cc3c22015-05-13 17:54:08 +00002648 testcase( pTerm->pExpr->op==TK_IS );
drh7a1bca72014-11-22 18:50:44 +00002649 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){
2650 k = 10;
2651 }else{
drhf06cdde2020-02-24 16:46:08 +00002652 k = 20;
drh7a1bca72014-11-22 18:50:44 +00002653 }
drh89efac92020-02-22 16:58:49 +00002654 if( iReduce<k ){
2655 pTerm->wtFlags |= TERM_HEURTRUTH;
2656 iReduce = k;
2657 }
drh7a1bca72014-11-22 18:50:44 +00002658 }
drhd8b77e22014-09-06 01:35:57 +00002659 }
danaa9933c2014-04-24 20:04:49 +00002660 }
drhcca9f3d2013-09-06 15:23:29 +00002661 }
drhfb82caf2021-12-08 19:50:45 +00002662 if( pLoop->nOut > nRow-iReduce ){
2663 pLoop->nOut = nRow - iReduce;
2664 }
drhcca9f3d2013-09-06 15:23:29 +00002665}
2666
dan71c57db2016-07-09 20:23:55 +00002667/*
2668** Term pTerm is a vector range comparison operation. The first comparison
dan553168c2016-08-01 20:14:31 +00002669** in the vector can be optimized using column nEq of the index. This
2670** function returns the total number of vector elements that can be used
2671** as part of the range comparison.
2672**
2673** For example, if the query is:
2674**
2675** WHERE a = ? AND (b, c, d) > (?, ?, ?)
2676**
2677** and the index:
2678**
2679** CREATE INDEX ... ON (a, b, c, d, e)
2680**
2681** then this function would be invoked with nEq=1. The value returned in
2682** this case is 3.
dan71c57db2016-07-09 20:23:55 +00002683*/
dan320d4c32016-10-08 11:55:12 +00002684static int whereRangeVectorLen(
drh64bcb8c2016-08-26 03:42:57 +00002685 Parse *pParse, /* Parsing context */
2686 int iCur, /* Cursor open on pIdx */
2687 Index *pIdx, /* The index to be used for a inequality constraint */
2688 int nEq, /* Number of prior equality constraints on same index */
2689 WhereTerm *pTerm /* The vector inequality constraint */
dan71c57db2016-07-09 20:23:55 +00002690){
2691 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft);
2692 int i;
2693
2694 nCmp = MIN(nCmp, (pIdx->nColumn - nEq));
2695 for(i=1; i<nCmp; i++){
2696 /* Test if comparison i of pTerm is compatible with column (i+nEq)
2697 ** of the index. If not, exit the loop. */
2698 char aff; /* Comparison affinity */
2699 char idxaff = 0; /* Indexed columns affinity */
2700 CollSeq *pColl; /* Comparison collation sequence */
drha4eeccd2021-10-07 17:43:30 +00002701 Expr *pLhs, *pRhs;
2702
2703 assert( ExprUseXList(pTerm->pExpr->pLeft) );
2704 pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr;
2705 pRhs = pTerm->pExpr->pRight;
2706 if( ExprUseXSelect(pRhs) ){
dan71c57db2016-07-09 20:23:55 +00002707 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr;
2708 }else{
2709 pRhs = pRhs->x.pList->a[i].pExpr;
2710 }
2711
2712 /* Check that the LHS of the comparison is a column reference to
dand05a7142016-08-02 17:07:51 +00002713 ** the right column of the right source table. And that the sort
2714 ** order of the index column is the same as the sort order of the
2715 ** leftmost index column. */
dan71c57db2016-07-09 20:23:55 +00002716 if( pLhs->op!=TK_COLUMN
2717 || pLhs->iTable!=iCur
2718 || pLhs->iColumn!=pIdx->aiColumn[i+nEq]
dan2c628ea2016-08-03 16:39:04 +00002719 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq]
dan71c57db2016-07-09 20:23:55 +00002720 ){
2721 break;
2722 }
2723
drh0c36fca2016-08-26 18:17:08 +00002724 testcase( pLhs->iColumn==XN_ROWID );
dan71c57db2016-07-09 20:23:55 +00002725 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs));
drh0dfa4f62016-08-26 13:19:49 +00002726 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn);
dan71c57db2016-07-09 20:23:55 +00002727 if( aff!=idxaff ) break;
2728
2729 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs);
drh0dfa4f62016-08-26 13:19:49 +00002730 if( pColl==0 ) break;
drh64bcb8c2016-08-26 03:42:57 +00002731 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break;
dan71c57db2016-07-09 20:23:55 +00002732 }
2733 return i;
2734}
2735
drhcca9f3d2013-09-06 15:23:29 +00002736/*
drhdbd94862014-07-23 23:57:42 +00002737** Adjust the cost C by the costMult facter T. This only occurs if
2738** compiled with -DSQLITE_ENABLE_COSTMULT
2739*/
2740#ifdef SQLITE_ENABLE_COSTMULT
2741# define ApplyCostMultiplier(C,T) C += T
2742#else
2743# define ApplyCostMultiplier(C,T)
2744#endif
2745
2746/*
dan4a6b8a02014-04-30 14:47:01 +00002747** We have so far matched pBuilder->pNew->u.btree.nEq terms of the
2748** index pIndex. Try to match one more.
2749**
2750** When this function is called, pBuilder->pNew->nOut contains the
2751** number of rows expected to be visited by filtering using the nEq
2752** terms only. If it is modified, this value is restored before this
2753** function returns.
drh1c8148f2013-05-04 20:25:23 +00002754**
drh5f913ec2019-01-10 13:56:08 +00002755** If pProbe->idxType==SQLITE_IDXTYPE_IPK, that means pIndex is
2756** a fake index used for the INTEGER PRIMARY KEY.
drh1c8148f2013-05-04 20:25:23 +00002757*/
drh5346e952013-05-08 14:14:26 +00002758static int whereLoopAddBtreeIndex(
drh1c8148f2013-05-04 20:25:23 +00002759 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */
drh76012942021-02-21 21:04:54 +00002760 SrcItem *pSrc, /* FROM clause term being analyzed */
drh1c8148f2013-05-04 20:25:23 +00002761 Index *pProbe, /* An index on pSrc */
drhbf539c42013-10-05 18:16:02 +00002762 LogEst nInMul /* log(Number of iterations due to IN) */
drh1c8148f2013-05-04 20:25:23 +00002763){
drh70d18342013-06-06 19:16:33 +00002764 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */
2765 Parse *pParse = pWInfo->pParse; /* Parsing context */
2766 sqlite3 *db = pParse->db; /* Database connection malloc context */
drh1c8148f2013-05-04 20:25:23 +00002767 WhereLoop *pNew; /* Template WhereLoop under construction */
2768 WhereTerm *pTerm; /* A WhereTerm under consideration */
drh43fe25f2013-05-07 23:06:23 +00002769 int opMask; /* Valid operators for constraints */
drh1c8148f2013-05-04 20:25:23 +00002770 WhereScan scan; /* Iterator for WHERE terms */
drh4efc9292013-06-06 23:02:03 +00002771 Bitmask saved_prereq; /* Original value of pNew->prereq */
2772 u16 saved_nLTerm; /* Original value of pNew->nLTerm */
drhcd8629e2013-11-13 12:27:25 +00002773 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */
dan71c57db2016-07-09 20:23:55 +00002774 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */
2775 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */
drhc8bbce12014-10-21 01:05:09 +00002776 u16 saved_nSkip; /* Original value of pNew->nSkip */
drh4efc9292013-06-06 23:02:03 +00002777 u32 saved_wsFlags; /* Original value of pNew->wsFlags */
drhbf539c42013-10-05 18:16:02 +00002778 LogEst saved_nOut; /* Original value of pNew->nOut */
drh5346e952013-05-08 14:14:26 +00002779 int rc = SQLITE_OK; /* Return code */
drhd8b77e22014-09-06 01:35:57 +00002780 LogEst rSize; /* Number of rows in the table */
drhbf539c42013-10-05 18:16:02 +00002781 LogEst rLogSize; /* Logarithm of table size */
drhc7f0d222013-06-19 03:27:12 +00002782 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */
drh1c8148f2013-05-04 20:25:23 +00002783
drh1c8148f2013-05-04 20:25:23 +00002784 pNew = pBuilder->pNew;
mistachkinfad30392016-02-13 23:43:46 +00002785 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT;
drhb035b872020-11-12 18:16:01 +00002786 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d, nSkip=%d, rRun=%d\n",
drhb592d472020-01-06 17:33:09 +00002787 pProbe->pTable->zName,pProbe->zName,
drhb035b872020-11-12 18:16:01 +00002788 pNew->u.btree.nEq, pNew->nSkip, pNew->rRun));
drh1c8148f2013-05-04 20:25:23 +00002789
drh5346e952013-05-08 14:14:26 +00002790 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 );
drh43fe25f2013-05-07 23:06:23 +00002791 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 );
2792 if( pNew->wsFlags & WHERE_BTM_LIMIT ){
2793 opMask = WO_LT|WO_LE;
drh1c8148f2013-05-04 20:25:23 +00002794 }else{
dan71c57db2016-07-09 20:23:55 +00002795 assert( pNew->u.btree.nBtm==0 );
drhe8d0c612015-05-14 01:05:25 +00002796 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS;
drh1c8148f2013-05-04 20:25:23 +00002797 }
drhef866372013-05-22 20:49:02 +00002798 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE);
drh1c8148f2013-05-04 20:25:23 +00002799
dan39129ce2014-06-30 15:23:57 +00002800 assert( pNew->u.btree.nEq<pProbe->nColumn );
drh756748e2021-05-13 13:43:40 +00002801 assert( pNew->u.btree.nEq<pProbe->nKeyCol
2802 || pProbe->idxType!=SQLITE_IDXTYPE_PRIMARYKEY );
dan39129ce2014-06-30 15:23:57 +00002803
drh4efc9292013-06-06 23:02:03 +00002804 saved_nEq = pNew->u.btree.nEq;
dan71c57db2016-07-09 20:23:55 +00002805 saved_nBtm = pNew->u.btree.nBtm;
2806 saved_nTop = pNew->u.btree.nTop;
drhc8bbce12014-10-21 01:05:09 +00002807 saved_nSkip = pNew->nSkip;
drh4efc9292013-06-06 23:02:03 +00002808 saved_nLTerm = pNew->nLTerm;
2809 saved_wsFlags = pNew->wsFlags;
2810 saved_prereq = pNew->prereq;
2811 saved_nOut = pNew->nOut;
drhbb523082015-08-27 15:58:51 +00002812 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq,
2813 opMask, pProbe);
drhb8a8e8a2013-06-10 19:12:39 +00002814 pNew->rSetup = 0;
drhd8b77e22014-09-06 01:35:57 +00002815 rSize = pProbe->aiRowLogEst[0];
2816 rLogSize = estLog(rSize);
drh5346e952013-05-08 14:14:26 +00002817 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){
dan8ad1d8b2014-04-25 20:22:45 +00002818 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */
danaa9933c2014-04-24 20:04:49 +00002819 LogEst rCostIdx;
dan8ad1d8b2014-04-25 20:22:45 +00002820 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */
drhb8a8e8a2013-06-10 19:12:39 +00002821 int nIn = 0;
drh175b8f02019-08-08 15:24:17 +00002822#ifdef SQLITE_ENABLE_STAT4
dan7a419232013-08-06 20:01:43 +00002823 int nRecValid = pBuilder->nRecValid;
drhb5246e52013-07-08 21:12:57 +00002824#endif
dan8ad1d8b2014-04-25 20:22:45 +00002825 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0)
drhbb523082015-08-27 15:58:51 +00002826 && indexColumnNotNull(pProbe, saved_nEq)
dan8bff07a2013-08-29 14:56:14 +00002827 ){
2828 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */
2829 }
dan7a419232013-08-06 20:01:43 +00002830 if( pTerm->prereqRight & pNew->maskSelf ) continue;
2831
drha40da622015-03-09 12:11:56 +00002832 /* Do not allow the upper bound of a LIKE optimization range constraint
2833 ** to mix with a lower range bound from some other source */
2834 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue;
2835
drhd4dae752019-12-11 16:22:53 +00002836 /* tag-20191211-001: Do not allow constraints from the WHERE clause to
drha76ac882022-04-08 19:20:12 +00002837 ** be used by the right table of a LEFT JOIN nor by the left table of a
2838 ** RIGHT JOIN. Only constraints in the
drhd4dae752019-12-11 16:22:53 +00002839 ** ON clause are allowed. See tag-20191211-002 for the vtab equivalent. */
drha76ac882022-04-08 19:20:12 +00002840 if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ))!=0
drh5996a772016-03-31 20:40:28 +00002841 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin)
drh5996a772016-03-31 20:40:28 +00002842 ){
drh5996a772016-03-31 20:40:28 +00002843 continue;
2844 }
2845
drha3928dd2017-02-17 15:26:36 +00002846 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){
drh89efac92020-02-22 16:58:49 +00002847 pBuilder->bldFlags1 |= SQLITE_BLDF1_UNIQUE;
drha3928dd2017-02-17 15:26:36 +00002848 }else{
drh89efac92020-02-22 16:58:49 +00002849 pBuilder->bldFlags1 |= SQLITE_BLDF1_INDEXED;
drha3928dd2017-02-17 15:26:36 +00002850 }
drh4efc9292013-06-06 23:02:03 +00002851 pNew->wsFlags = saved_wsFlags;
2852 pNew->u.btree.nEq = saved_nEq;
dan71c57db2016-07-09 20:23:55 +00002853 pNew->u.btree.nBtm = saved_nBtm;
2854 pNew->u.btree.nTop = saved_nTop;
drh4efc9292013-06-06 23:02:03 +00002855 pNew->nLTerm = saved_nLTerm;
2856 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */
2857 pNew->aLTerm[pNew->nLTerm++] = pTerm;
2858 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf;
dan8ad1d8b2014-04-25 20:22:45 +00002859
2860 assert( nInMul==0
2861 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0
2862 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0
2863 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0
2864 );
2865
2866 if( eOp & WO_IN ){
drha18f3d22013-05-08 03:05:41 +00002867 Expr *pExpr = pTerm->pExpr;
drha4eeccd2021-10-07 17:43:30 +00002868 if( ExprUseXSelect(pExpr) ){
drhe1e2e9a2013-06-13 15:16:53 +00002869 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */
dan3d1fb1d2016-08-03 18:00:49 +00002870 int i;
drhbf539c42013-10-05 18:16:02 +00002871 nIn = 46; assert( 46==sqlite3LogEst(25) );
dan3d1fb1d2016-08-03 18:00:49 +00002872
2873 /* The expression may actually be of the form (x, y) IN (SELECT...).
2874 ** In this case there is a separate term for each of (x) and (y).
2875 ** However, the nIn multiplier should only be applied once, not once
2876 ** for each such term. The following loop checks that pTerm is the
2877 ** first such term in use, and sets nIn back to 0 if it is not. */
2878 for(i=0; i<pNew->nLTerm-1; i++){
drh9a2e5162016-09-19 11:00:42 +00002879 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0;
dan3d1fb1d2016-08-03 18:00:49 +00002880 }
drha18f3d22013-05-08 03:05:41 +00002881 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){
2882 /* "x IN (value, value, ...)" */
drhbf539c42013-10-05 18:16:02 +00002883 nIn = sqlite3LogEst(pExpr->x.pList->nExpr);
drhf1645f02013-05-07 19:44:38 +00002884 }
drh7d14ffe2020-10-02 13:48:57 +00002885 if( pProbe->hasStat1 && rLogSize>=10 ){
drhb034a242021-06-02 12:44:26 +00002886 LogEst M, logK, x;
drh6d6decb2018-06-08 21:21:01 +00002887 /* Let:
2888 ** N = the total number of rows in the table
2889 ** K = the number of entries on the RHS of the IN operator
2890 ** M = the number of rows in the table that match terms to the
2891 ** to the left in the same index. If the IN operator is on
2892 ** the left-most index column, M==N.
2893 **
2894 ** Given the definitions above, it is better to omit the IN operator
2895 ** from the index lookup and instead do a scan of the M elements,
2896 ** testing each scanned row against the IN operator separately, if:
2897 **
2898 ** M*log(K) < K*log(N)
2899 **
2900 ** Our estimates for M, K, and N might be inaccurate, so we build in
2901 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator
2902 ** with the index, as using an index has better worst-case behavior.
2903 ** If we do not have real sqlite_stat1 data, always prefer to use
drh7d14ffe2020-10-02 13:48:57 +00002904 ** the index. Do not bother with this optimization on very small
2905 ** tables (less than 2 rows) as it is pointless in that case.
drh6d6decb2018-06-08 21:21:01 +00002906 */
2907 M = pProbe->aiRowLogEst[saved_nEq];
2908 logK = estLog(nIn);
drhb034a242021-06-02 12:44:26 +00002909 /* TUNING v----- 10 to bias toward indexed IN */
2910 x = M + logK + 10 - (nIn + rLogSize);
2911 if( x>=0 ){
drh6d6decb2018-06-08 21:21:01 +00002912 WHERETRACE(0x40,
drhb034a242021-06-02 12:44:26 +00002913 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d) "
2914 "prefers indexed lookup\n",
2915 saved_nEq, M, logK, nIn, rLogSize, x));
drh3074faa2021-06-02 19:28:07 +00002916 }else if( nInMul<2 && OptimizationEnabled(db, SQLITE_SeekScan) ){
drhb034a242021-06-02 12:44:26 +00002917 WHERETRACE(0x40,
2918 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d"
2919 " nInMul=%d) prefers skip-scan\n",
2920 saved_nEq, M, logK, nIn, rLogSize, x, nInMul));
drh68cf0ac2020-09-28 19:51:54 +00002921 pNew->wsFlags |= WHERE_IN_SEEKSCAN;
drh6d6decb2018-06-08 21:21:01 +00002922 }else{
2923 WHERETRACE(0x40,
drhb034a242021-06-02 12:44:26 +00002924 ("IN operator (N=%d M=%d logK=%d nIn=%d rLogSize=%d x=%d"
2925 " nInMul=%d) prefers normal scan\n",
2926 saved_nEq, M, logK, nIn, rLogSize, x, nInMul));
2927 continue;
drh6d6decb2018-06-08 21:21:01 +00002928 }
drhda4c4092018-06-08 18:22:10 +00002929 }
2930 pNew->wsFlags |= WHERE_COLUMN_IN;
drhe8d0c612015-05-14 01:05:25 +00002931 }else if( eOp & (WO_EQ|WO_IS) ){
drhbb523082015-08-27 15:58:51 +00002932 int iCol = pProbe->aiColumn[saved_nEq];
drha18f3d22013-05-08 03:05:41 +00002933 pNew->wsFlags |= WHERE_COLUMN_EQ;
drhbb523082015-08-27 15:58:51 +00002934 assert( saved_nEq==pNew->u.btree.nEq );
drh4b92f982015-09-29 17:20:14 +00002935 if( iCol==XN_ROWID
dan75dbf682017-11-20 14:40:03 +00002936 || (iCol>=0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1)
drh4b92f982015-09-29 17:20:14 +00002937 ){
dan4ea48142018-01-31 14:07:01 +00002938 if( iCol==XN_ROWID || pProbe->uniqNotNull
dan8433e712018-01-29 17:08:52 +00002939 || (pProbe->nKeyCol==1 && pProbe->onError && eOp==WO_EQ)
2940 ){
drhe39a7322014-02-03 14:04:11 +00002941 pNew->wsFlags |= WHERE_ONEROW;
dan8433e712018-01-29 17:08:52 +00002942 }else{
2943 pNew->wsFlags |= WHERE_UNQ_WANTED;
drhe39a7322014-02-03 14:04:11 +00002944 }
drh21f7ff72013-06-03 15:07:23 +00002945 }
drh67656ac2021-05-04 23:21:35 +00002946 if( scan.iEquiv>1 ) pNew->wsFlags |= WHERE_TRANSCONS;
dan2dd3cdc2014-04-26 20:21:14 +00002947 }else if( eOp & WO_ISNULL ){
2948 pNew->wsFlags |= WHERE_COLUMN_NULL;
dan8ad1d8b2014-04-25 20:22:45 +00002949 }else if( eOp & (WO_GT|WO_GE) ){
2950 testcase( eOp & WO_GT );
2951 testcase( eOp & WO_GE );
drha18f3d22013-05-08 03:05:41 +00002952 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT;
dan71c57db2016-07-09 20:23:55 +00002953 pNew->u.btree.nBtm = whereRangeVectorLen(
2954 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm
2955 );
drh6f2bfad2013-06-03 17:35:22 +00002956 pBtm = pTerm;
2957 pTop = 0;
drha40da622015-03-09 12:11:56 +00002958 if( pTerm->wtFlags & TERM_LIKEOPT ){
mistachkin9a60e712020-12-22 19:57:53 +00002959 /* Range constraints that come from the LIKE optimization are
drh80314622015-03-09 13:01:02 +00002960 ** always used in pairs. */
drha40da622015-03-09 12:11:56 +00002961 pTop = &pTerm[1];
2962 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm );
2963 assert( pTop->wtFlags & TERM_LIKEOPT );
2964 assert( pTop->eOperator==WO_LT );
2965 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */
2966 pNew->aLTerm[pNew->nLTerm++] = pTop;
2967 pNew->wsFlags |= WHERE_TOP_LIMIT;
dan71c57db2016-07-09 20:23:55 +00002968 pNew->u.btree.nTop = 1;
drha40da622015-03-09 12:11:56 +00002969 }
dan2dd3cdc2014-04-26 20:21:14 +00002970 }else{
dan8ad1d8b2014-04-25 20:22:45 +00002971 assert( eOp & (WO_LT|WO_LE) );
2972 testcase( eOp & WO_LT );
2973 testcase( eOp & WO_LE );
drha18f3d22013-05-08 03:05:41 +00002974 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT;
dan71c57db2016-07-09 20:23:55 +00002975 pNew->u.btree.nTop = whereRangeVectorLen(
2976 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm
2977 );
drh6f2bfad2013-06-03 17:35:22 +00002978 pTop = pTerm;
2979 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ?
drh4efc9292013-06-06 23:02:03 +00002980 pNew->aLTerm[pNew->nLTerm-2] : 0;
drh1c8148f2013-05-04 20:25:23 +00002981 }
dan8ad1d8b2014-04-25 20:22:45 +00002982
2983 /* At this point pNew->nOut is set to the number of rows expected to
2984 ** be visited by the index scan before considering term pTerm, or the
2985 ** values of nIn and nInMul. In other words, assuming that all
2986 ** "x IN(...)" terms are replaced with "x = ?". This block updates
2987 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */
2988 assert( pNew->nOut==saved_nOut );
drh6f2bfad2013-06-03 17:35:22 +00002989 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){
drh175b8f02019-08-08 15:24:17 +00002990 /* Adjust nOut using stat4 data. Or, if there is no stat4
danaa9933c2014-04-24 20:04:49 +00002991 ** data, using some other estimate. */
drh186ad8c2013-10-08 18:40:37 +00002992 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew);
dan8ad1d8b2014-04-25 20:22:45 +00002993 }else{
2994 int nEq = ++pNew->u.btree.nEq;
drhe8d0c612015-05-14 01:05:25 +00002995 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) );
dan8ad1d8b2014-04-25 20:22:45 +00002996
2997 assert( pNew->nOut==saved_nOut );
drhbb523082015-08-27 15:58:51 +00002998 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){
dan8ad1d8b2014-04-25 20:22:45 +00002999 assert( (eOp & WO_IN) || nIn==0 );
drhc5f246e2014-05-01 20:24:21 +00003000 testcase( eOp & WO_IN );
dan8ad1d8b2014-04-25 20:22:45 +00003001 pNew->nOut += pTerm->truthProb;
3002 pNew->nOut -= nIn;
dan8ad1d8b2014-04-25 20:22:45 +00003003 }else{
drh175b8f02019-08-08 15:24:17 +00003004#ifdef SQLITE_ENABLE_STAT4
dan8ad1d8b2014-04-25 20:22:45 +00003005 tRowcnt nOut = 0;
3006 if( nInMul==0
3007 && pProbe->nSample
drh9c32c912021-06-16 19:23:24 +00003008 && ALWAYS(pNew->u.btree.nEq<=pProbe->nSampleCol)
drha4eeccd2021-10-07 17:43:30 +00003009 && ((eOp & WO_IN)==0 || ExprUseXList(pTerm->pExpr))
drh5eae1d12019-08-08 16:23:12 +00003010 && OptimizationEnabled(db, SQLITE_Stat4)
dan8ad1d8b2014-04-25 20:22:45 +00003011 ){
3012 Expr *pExpr = pTerm->pExpr;
drhe8d0c612015-05-14 01:05:25 +00003013 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){
3014 testcase( eOp & WO_EQ );
3015 testcase( eOp & WO_IS );
dan8ad1d8b2014-04-25 20:22:45 +00003016 testcase( eOp & WO_ISNULL );
3017 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut);
3018 }else{
3019 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut);
3020 }
dan8ad1d8b2014-04-25 20:22:45 +00003021 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK;
3022 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */
3023 if( nOut ){
3024 pNew->nOut = sqlite3LogEst(nOut);
drhcea19512020-02-22 18:27:48 +00003025 if( nEq==1
drhf06cdde2020-02-24 16:46:08 +00003026 /* TUNING: Mark terms as "low selectivity" if they seem likely
3027 ** to be true for half or more of the rows in the table.
3028 ** See tag-202002240-1 */
3029 && pNew->nOut+10 > pProbe->aiRowLogEst[0]
drhcea19512020-02-22 18:27:48 +00003030 ){
drh89efac92020-02-22 16:58:49 +00003031#if WHERETRACE_ENABLED /* 0x01 */
3032 if( sqlite3WhereTrace & 0x01 ){
drhf06cdde2020-02-24 16:46:08 +00003033 sqlite3DebugPrintf(
3034 "STAT4 determines term has low selectivity:\n");
drh89efac92020-02-22 16:58:49 +00003035 sqlite3WhereTermPrint(pTerm, 999);
3036 }
3037#endif
drhf06cdde2020-02-24 16:46:08 +00003038 pTerm->wtFlags |= TERM_HIGHTRUTH;
drh89efac92020-02-22 16:58:49 +00003039 if( pTerm->wtFlags & TERM_HEURTRUTH ){
drhf06cdde2020-02-24 16:46:08 +00003040 /* If the term has previously been used with an assumption of
3041 ** higher selectivity, then set the flag to rerun the
3042 ** loop computations. */
drh89efac92020-02-22 16:58:49 +00003043 pBuilder->bldFlags2 |= SQLITE_BLDF2_2NDPASS;
3044 }
3045 }
dan8ad1d8b2014-04-25 20:22:45 +00003046 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut;
3047 pNew->nOut -= nIn;
3048 }
3049 }
3050 if( nOut==0 )
3051#endif
3052 {
3053 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]);
3054 if( eOp & WO_ISNULL ){
3055 /* TUNING: If there is no likelihood() value, assume that a
3056 ** "col IS NULL" expression matches twice as many rows
3057 ** as (col=?). */
3058 pNew->nOut += 10;
3059 }
3060 }
dan6cb8d762013-08-08 11:48:57 +00003061 }
drh6f2bfad2013-06-03 17:35:22 +00003062 }
dan8ad1d8b2014-04-25 20:22:45 +00003063
danaa9933c2014-04-24 20:04:49 +00003064 /* Set rCostIdx to the cost of visiting selected rows in index. Add
3065 ** it to pNew->rRun, which is currently set to the cost of the index
3066 ** seek only. Then, if this is a non-covering index, add the cost of
3067 ** visiting the rows in the main table. */
drh725dd722019-08-15 14:35:45 +00003068 assert( pSrc->pTab->szTabRow>0 );
danaa9933c2014-04-24 20:04:49 +00003069 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow;
dan8ad1d8b2014-04-25 20:22:45 +00003070 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx);
drhe217efc2013-06-12 03:48:41 +00003071 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){
danaa9933c2014-04-24 20:04:49 +00003072 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16);
drheb04de32013-05-10 15:16:30 +00003073 }
drhdbd94862014-07-23 23:57:42 +00003074 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult);
danaa9933c2014-04-24 20:04:49 +00003075
dan8ad1d8b2014-04-25 20:22:45 +00003076 nOutUnadjusted = pNew->nOut;
3077 pNew->rRun += nInMul + nIn;
3078 pNew->nOut += nInMul + nIn;
drhd8b77e22014-09-06 01:35:57 +00003079 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize);
drhcf8fa7a2013-05-10 20:26:22 +00003080 rc = whereLoopInsert(pBuilder, pNew);
dan440e6ff2014-04-28 08:49:54 +00003081
3082 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){
3083 pNew->nOut = saved_nOut;
3084 }else{
3085 pNew->nOut = nOutUnadjusted;
3086 }
dan8ad1d8b2014-04-25 20:22:45 +00003087
drh5346e952013-05-08 14:14:26 +00003088 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0
dan39129ce2014-06-30 15:23:57 +00003089 && pNew->u.btree.nEq<pProbe->nColumn
drh756748e2021-05-13 13:43:40 +00003090 && (pNew->u.btree.nEq<pProbe->nKeyCol ||
3091 pProbe->idxType!=SQLITE_IDXTYPE_PRIMARYKEY)
drh5346e952013-05-08 14:14:26 +00003092 ){
drhb8a8e8a2013-06-10 19:12:39 +00003093 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn);
drha18f3d22013-05-08 03:05:41 +00003094 }
danad45ed72013-08-08 12:21:32 +00003095 pNew->nOut = saved_nOut;
drh175b8f02019-08-08 15:24:17 +00003096#ifdef SQLITE_ENABLE_STAT4
dan7a419232013-08-06 20:01:43 +00003097 pBuilder->nRecValid = nRecValid;
dan7a419232013-08-06 20:01:43 +00003098#endif
drh1c8148f2013-05-04 20:25:23 +00003099 }
drh4efc9292013-06-06 23:02:03 +00003100 pNew->prereq = saved_prereq;
3101 pNew->u.btree.nEq = saved_nEq;
dan71c57db2016-07-09 20:23:55 +00003102 pNew->u.btree.nBtm = saved_nBtm;
3103 pNew->u.btree.nTop = saved_nTop;
drhc8bbce12014-10-21 01:05:09 +00003104 pNew->nSkip = saved_nSkip;
drh4efc9292013-06-06 23:02:03 +00003105 pNew->wsFlags = saved_wsFlags;
3106 pNew->nOut = saved_nOut;
3107 pNew->nLTerm = saved_nLTerm;
drhc8bbce12014-10-21 01:05:09 +00003108
3109 /* Consider using a skip-scan if there are no WHERE clause constraints
3110 ** available for the left-most terms of the index, and if the average
3111 ** number of repeats in the left-most terms is at least 18.
3112 **
3113 ** The magic number 18 is selected on the basis that scanning 17 rows
3114 ** is almost always quicker than an index seek (even though if the index
3115 ** contains fewer than 2^17 rows we assume otherwise in other parts of
3116 ** the code). And, even if it is not, it should not be too much slower.
3117 ** On the other hand, the extra seeks could end up being significantly
3118 ** more expensive. */
3119 assert( 42==sqlite3LogEst(18) );
3120 if( saved_nEq==saved_nSkip
3121 && saved_nEq+1<pProbe->nKeyCol
drhb592d472020-01-06 17:33:09 +00003122 && saved_nEq==pNew->nLTerm
drhf9df2fb2014-11-15 19:08:13 +00003123 && pProbe->noSkipScan==0
drhab7fdca2020-02-13 14:51:54 +00003124 && pProbe->hasStat1!=0
dane8825512018-07-12 19:14:39 +00003125 && OptimizationEnabled(db, SQLITE_SkipScan)
drhc8bbce12014-10-21 01:05:09 +00003126 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */
3127 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK
3128 ){
3129 LogEst nIter;
3130 pNew->u.btree.nEq++;
3131 pNew->nSkip++;
3132 pNew->aLTerm[pNew->nLTerm++] = 0;
3133 pNew->wsFlags |= WHERE_SKIPSCAN;
3134 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1];
drhc8bbce12014-10-21 01:05:09 +00003135 pNew->nOut -= nIter;
3136 /* TUNING: Because uncertainties in the estimates for skip-scan queries,
3137 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */
3138 nIter += 5;
3139 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul);
3140 pNew->nOut = saved_nOut;
3141 pNew->u.btree.nEq = saved_nEq;
3142 pNew->nSkip = saved_nSkip;
3143 pNew->wsFlags = saved_wsFlags;
3144 }
3145
drh0f1631d2018-04-09 13:58:20 +00003146 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n",
3147 pProbe->pTable->zName, pProbe->zName, saved_nEq, rc));
drh5346e952013-05-08 14:14:26 +00003148 return rc;
drh1c8148f2013-05-04 20:25:23 +00003149}
3150
3151/*
drh23f98da2013-05-21 15:52:07 +00003152** Return True if it is possible that pIndex might be useful in
3153** implementing the ORDER BY clause in pBuilder.
3154**
3155** Return False if pBuilder does not contain an ORDER BY clause or
3156** if there is no way for pIndex to be useful in implementing that
3157** ORDER BY clause.
3158*/
3159static int indexMightHelpWithOrderBy(
3160 WhereLoopBuilder *pBuilder,
3161 Index *pIndex,
3162 int iCursor
3163){
3164 ExprList *pOB;
drhdae26fe2015-09-24 18:47:59 +00003165 ExprList *aColExpr;
drh6d381472013-06-13 17:58:08 +00003166 int ii, jj;
drh23f98da2013-05-21 15:52:07 +00003167
drh53cfbe92013-06-13 17:28:22 +00003168 if( pIndex->bUnordered ) return 0;
drh70d18342013-06-06 19:16:33 +00003169 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0;
drh23f98da2013-05-21 15:52:07 +00003170 for(ii=0; ii<pOB->nExpr; ii++){
drh0d950af2019-08-22 16:38:42 +00003171 Expr *pExpr = sqlite3ExprSkipCollateAndLikely(pOB->a[ii].pExpr);
drh235667a2020-11-08 20:44:30 +00003172 if( NEVER(pExpr==0) ) continue;
drhdae26fe2015-09-24 18:47:59 +00003173 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){
drh137fd4f2014-09-19 02:01:37 +00003174 if( pExpr->iColumn<0 ) return 1;
drhbbbdc832013-10-22 18:01:40 +00003175 for(jj=0; jj<pIndex->nKeyCol; jj++){
drh6d381472013-06-13 17:58:08 +00003176 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1;
3177 }
drhdae26fe2015-09-24 18:47:59 +00003178 }else if( (aColExpr = pIndex->aColExpr)!=0 ){
3179 for(jj=0; jj<pIndex->nKeyCol; jj++){
drh4b92f982015-09-29 17:20:14 +00003180 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue;
drhdb8e68b2017-09-28 01:09:42 +00003181 if( sqlite3ExprCompareSkip(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){
drhdae26fe2015-09-24 18:47:59 +00003182 return 1;
3183 }
3184 }
drh23f98da2013-05-21 15:52:07 +00003185 }
3186 }
3187 return 0;
3188}
3189
drh4bd5f732013-07-31 23:22:39 +00003190/* Check to see if a partial index with pPartIndexWhere can be used
3191** in the current query. Return true if it can be and false if not.
3192*/
drhca7a26b2019-11-30 19:29:19 +00003193static int whereUsablePartialIndex(
3194 int iTab, /* The table for which we want an index */
3195 int isLeft, /* True if iTab is the right table of a LEFT JOIN */
3196 WhereClause *pWC, /* The WHERE clause of the query */
3197 Expr *pWhere /* The WHERE clause from the partial index */
3198){
drh4bd5f732013-07-31 23:22:39 +00003199 int i;
3200 WhereTerm *pTerm;
drh3e380a42017-06-28 18:25:03 +00003201 Parse *pParse = pWC->pWInfo->pParse;
drhcf599b62015-08-07 20:57:00 +00003202 while( pWhere->op==TK_AND ){
drhca7a26b2019-11-30 19:29:19 +00003203 if( !whereUsablePartialIndex(iTab,isLeft,pWC,pWhere->pLeft) ) return 0;
drhcf599b62015-08-07 20:57:00 +00003204 pWhere = pWhere->pRight;
3205 }
drh3e380a42017-06-28 18:25:03 +00003206 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0;
drh4bd5f732013-07-31 23:22:39 +00003207 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
drhc7d12f42019-09-03 14:27:25 +00003208 Expr *pExpr;
drhc7d12f42019-09-03 14:27:25 +00003209 pExpr = pTerm->pExpr;
drhd1985262022-04-11 11:25:28 +00003210 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->w.iJoin==iTab)
drhca7a26b2019-11-30 19:29:19 +00003211 && (isLeft==0 || ExprHasProperty(pExpr, EP_FromJoin))
drhd65b7e32021-05-29 23:07:59 +00003212 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab)
3213 && (pTerm->wtFlags & TERM_VNULL)==0
drh077f06e2015-02-24 16:48:59 +00003214 ){
3215 return 1;
3216 }
drh4bd5f732013-07-31 23:22:39 +00003217 }
3218 return 0;
3219}
drh92a121f2013-06-10 12:15:47 +00003220
3221/*
dan51576f42013-07-02 10:06:15 +00003222** Add all WhereLoop objects for a single table of the join where the table
dan71c57db2016-07-09 20:23:55 +00003223** is identified by pBuilder->pNew->iTab. That table is guaranteed to be
drh0823c892013-05-11 00:06:23 +00003224** a b-tree table, not a virtual table.
dan81647222014-04-30 15:00:16 +00003225**
3226** The costs (WhereLoop.rRun) of the b-tree loops added by this function
3227** are calculated as follows:
3228**
3229** For a full scan, assuming the table (or index) contains nRow rows:
3230**
3231** cost = nRow * 3.0 // full-table scan
3232** cost = nRow * K // scan of covering index
3233** cost = nRow * (K+3.0) // scan of non-covering index
3234**
3235** where K is a value between 1.1 and 3.0 set based on the relative
3236** estimated average size of the index and table records.
3237**
3238** For an index scan, where nVisit is the number of index rows visited
3239** by the scan, and nSeek is the number of seek operations required on
3240** the index b-tree:
3241**
3242** cost = nSeek * (log(nRow) + K * nVisit) // covering index
3243** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index
3244**
3245** Normally, nSeek is 1. nSeek values greater than 1 come about if the
3246** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when
3247** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans.
drh83a305f2014-07-22 12:05:32 +00003248**
3249** The estimated values (nRow, nVisit, nSeek) often contain a large amount
3250** of uncertainty. For this reason, scoring is designed to pick plans that
3251** "do the least harm" if the estimates are inaccurate. For example, a
3252** log(nRow) factor is omitted from a non-covering index scan in order to
3253** bias the scoring in favor of using an index, since the worst-case
3254** performance of using an index is far better than the worst-case performance
3255** of a full table scan.
drhf1b5f5b2013-05-02 00:15:01 +00003256*/
drh5346e952013-05-08 14:14:26 +00003257static int whereLoopAddBtree(
drh1c8148f2013-05-04 20:25:23 +00003258 WhereLoopBuilder *pBuilder, /* WHERE clause information */
drh599d5762016-03-08 01:11:51 +00003259 Bitmask mPrereq /* Extra prerequesites for using this table */
drhf1b5f5b2013-05-02 00:15:01 +00003260){
drh70d18342013-06-06 19:16:33 +00003261 WhereInfo *pWInfo; /* WHERE analysis context */
drh1c8148f2013-05-04 20:25:23 +00003262 Index *pProbe; /* An index we are evaluating */
drh1c8148f2013-05-04 20:25:23 +00003263 Index sPk; /* A fake index object for the primary key */
dancfc9df72014-04-25 15:01:01 +00003264 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */
drhbbbdc832013-10-22 18:01:40 +00003265 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */
drh70d18342013-06-06 19:16:33 +00003266 SrcList *pTabList; /* The FROM clause */
drh76012942021-02-21 21:04:54 +00003267 SrcItem *pSrc; /* The FROM clause btree term to add */
drh1c8148f2013-05-04 20:25:23 +00003268 WhereLoop *pNew; /* Template WhereLoop object */
drh5346e952013-05-08 14:14:26 +00003269 int rc = SQLITE_OK; /* Return code */
drhd044d202013-05-31 12:43:55 +00003270 int iSortIdx = 1; /* Index number */
drh23f98da2013-05-21 15:52:07 +00003271 int b; /* A boolean value */
drhbf539c42013-10-05 18:16:02 +00003272 LogEst rSize; /* number of rows in the table */
drh4bd5f732013-07-31 23:22:39 +00003273 WhereClause *pWC; /* The parsed WHERE clause */
drh3495d202013-10-07 17:32:15 +00003274 Table *pTab; /* Table being queried */
drh23f98da2013-05-21 15:52:07 +00003275
drh1c8148f2013-05-04 20:25:23 +00003276 pNew = pBuilder->pNew;
drh70d18342013-06-06 19:16:33 +00003277 pWInfo = pBuilder->pWInfo;
3278 pTabList = pWInfo->pTabList;
3279 pSrc = pTabList->a + pNew->iTab;
drh3495d202013-10-07 17:32:15 +00003280 pTab = pSrc->pTab;
drh4bd5f732013-07-31 23:22:39 +00003281 pWC = pBuilder->pWC;
drh0823c892013-05-11 00:06:23 +00003282 assert( !IsVirtual(pSrc->pTab) );
drh1c8148f2013-05-04 20:25:23 +00003283
drh271d7c22021-02-20 13:36:14 +00003284 if( pSrc->fg.isIndexedBy ){
drhdbfbb5a2021-10-07 23:04:50 +00003285 assert( pSrc->fg.isCte==0 );
drh1c8148f2013-05-04 20:25:23 +00003286 /* An INDEXED BY clause specifies a particular index to use */
drha79e2a22021-02-21 23:44:14 +00003287 pProbe = pSrc->u2.pIBIndex;
drhec95c442013-10-23 01:57:32 +00003288 }else if( !HasRowid(pTab) ){
3289 pProbe = pTab->pIndex;
drh1c8148f2013-05-04 20:25:23 +00003290 }else{
3291 /* There is no INDEXED BY clause. Create a fake Index object in local
3292 ** variable sPk to represent the rowid primary key index. Make this
3293 ** fake index the first in a chain of Index objects with all of the real
3294 ** indices to follow */
3295 Index *pFirst; /* First of real indices on the table */
3296 memset(&sPk, 0, sizeof(Index));
drhbbbdc832013-10-22 18:01:40 +00003297 sPk.nKeyCol = 1;
dan39129ce2014-06-30 15:23:57 +00003298 sPk.nColumn = 1;
drh1c8148f2013-05-04 20:25:23 +00003299 sPk.aiColumn = &aiColumnPk;
dancfc9df72014-04-25 15:01:01 +00003300 sPk.aiRowLogEst = aiRowEstPk;
drh1c8148f2013-05-04 20:25:23 +00003301 sPk.onError = OE_Replace;
drh3495d202013-10-07 17:32:15 +00003302 sPk.pTable = pTab;
danaa9933c2014-04-24 20:04:49 +00003303 sPk.szIdxRow = pTab->szTabRow;
drh5f913ec2019-01-10 13:56:08 +00003304 sPk.idxType = SQLITE_IDXTYPE_IPK;
dancfc9df72014-04-25 15:01:01 +00003305 aiRowEstPk[0] = pTab->nRowLogEst;
3306 aiRowEstPk[1] = 0;
drh1c8148f2013-05-04 20:25:23 +00003307 pFirst = pSrc->pTab->pIndex;
drh8a48b9c2015-08-19 15:20:00 +00003308 if( pSrc->fg.notIndexed==0 ){
drh1c8148f2013-05-04 20:25:23 +00003309 /* The real indices of the table are only considered if the
3310 ** NOT INDEXED qualifier is omitted from the FROM clause */
3311 sPk.pNext = pFirst;
3312 }
3313 pProbe = &sPk;
3314 }
dancfc9df72014-04-25 15:01:01 +00003315 rSize = pTab->nRowLogEst;
drheb04de32013-05-10 15:16:30 +00003316
drhfeb56e02013-08-23 17:33:46 +00003317#ifndef SQLITE_OMIT_AUTOMATIC_INDEX
drheb04de32013-05-10 15:16:30 +00003318 /* Automatic indexes */
drh8a48b9c2015-08-19 15:20:00 +00003319 if( !pBuilder->pOrSet /* Not part of an OR optimization */
drhd345dcf2022-04-14 14:58:50 +00003320 && (pWInfo->wctrlFlags & (WHERE_RIGHT_JOIN|WHERE_OR_SUBCLAUSE))==0
drh4fe425a2013-06-12 17:08:06 +00003321 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0
drh271d7c22021-02-20 13:36:14 +00003322 && !pSrc->fg.isIndexedBy /* Has no INDEXED BY clause */
drh8a48b9c2015-08-19 15:20:00 +00003323 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */
drh76226dd2015-09-24 17:38:01 +00003324 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */
drh8a48b9c2015-08-19 15:20:00 +00003325 && !pSrc->fg.isCorrelated /* Not a correlated subquery */
3326 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */
drh529394e2022-04-10 23:48:47 +00003327 && (pSrc->fg.jointype & JT_RIGHT)==0 /* Not the right tab of a RIGHT JOIN */
drheb04de32013-05-10 15:16:30 +00003328 ){
3329 /* Generate auto-index WhereLoops */
drh9045e7d2021-07-28 01:22:23 +00003330 LogEst rLogSize; /* Logarithm of the number of rows in the table */
drheb04de32013-05-10 15:16:30 +00003331 WhereTerm *pTerm;
3332 WhereTerm *pWCEnd = pWC->a + pWC->nTerm;
drh9045e7d2021-07-28 01:22:23 +00003333 rLogSize = estLog(rSize);
drheb04de32013-05-10 15:16:30 +00003334 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){
drh79a13bf2013-05-31 20:28:28 +00003335 if( pTerm->prereqRight & pNew->maskSelf ) continue;
drheb04de32013-05-10 15:16:30 +00003336 if( termCanDriveIndex(pTerm, pSrc, 0) ){
3337 pNew->u.btree.nEq = 1;
drhc8bbce12014-10-21 01:05:09 +00003338 pNew->nSkip = 0;
drhef866372013-05-22 20:49:02 +00003339 pNew->u.btree.pIndex = 0;
drh4efc9292013-06-06 23:02:03 +00003340 pNew->nLTerm = 1;
3341 pNew->aLTerm[0] = pTerm;
drhe1e2e9a2013-06-13 15:16:53 +00003342 /* TUNING: One-time cost for computing the automatic index is
drh7e074332014-09-22 14:30:51 +00003343 ** estimated to be X*N*log2(N) where N is the number of rows in
3344 ** the table being indexed and where X is 7 (LogEst=28) for normal
drh492ad132018-05-14 22:46:11 +00003345 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value
drh7e074332014-09-22 14:30:51 +00003346 ** of X is smaller for views and subqueries so that the query planner
3347 ** will be more aggressive about generating automatic indexes for
3348 ** those objects, since there is no opportunity to add schema
3349 ** indexes on subqueries and views. */
drh492ad132018-05-14 22:46:11 +00003350 pNew->rSetup = rLogSize + rSize;
drhf38524d2021-08-02 16:41:57 +00003351 if( !IsView(pTab) && (pTab->tabFlags & TF_Ephemeral)==0 ){
drh492ad132018-05-14 22:46:11 +00003352 pNew->rSetup += 28;
3353 }else{
3354 pNew->rSetup -= 10;
drh7e074332014-09-22 14:30:51 +00003355 }
drhdbd94862014-07-23 23:57:42 +00003356 ApplyCostMultiplier(pNew->rSetup, pTab->costMult);
drh05d1bad2016-06-26 04:06:28 +00003357 if( pNew->rSetup<0 ) pNew->rSetup = 0;
drh986b3872013-06-28 21:12:20 +00003358 /* TUNING: Each index lookup yields 20 rows in the table. This
3359 ** is more than the usual guess of 10 rows, since we have no way
peter.d.reid60ec9142014-09-06 16:39:46 +00003360 ** of knowing how selective the index will ultimately be. It would
drh986b3872013-06-28 21:12:20 +00003361 ** not be unreasonable to make this value much larger. */
drhbf539c42013-10-05 18:16:02 +00003362 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) );
drhb50596d2013-10-08 20:42:41 +00003363 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut);
drh986b3872013-06-28 21:12:20 +00003364 pNew->wsFlags = WHERE_AUTO_INDEX;
drh599d5762016-03-08 01:11:51 +00003365 pNew->prereq = mPrereq | pTerm->prereqRight;
drhcf8fa7a2013-05-10 20:26:22 +00003366 rc = whereLoopInsert(pBuilder, pNew);
drheb04de32013-05-10 15:16:30 +00003367 }
3368 }
3369 }
drhfeb56e02013-08-23 17:33:46 +00003370#endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
drh1c8148f2013-05-04 20:25:23 +00003371
dan85e1f462017-11-07 18:20:15 +00003372 /* Loop over all indices. If there was an INDEXED BY clause, then only
3373 ** consider index pProbe. */
3374 for(; rc==SQLITE_OK && pProbe;
drh271d7c22021-02-20 13:36:14 +00003375 pProbe=(pSrc->fg.isIndexedBy ? 0 : pProbe->pNext), iSortIdx++
dan85e1f462017-11-07 18:20:15 +00003376 ){
drhca7a26b2019-11-30 19:29:19 +00003377 int isLeft = (pSrc->fg.jointype & JT_OUTER)!=0;
drh4bd5f732013-07-31 23:22:39 +00003378 if( pProbe->pPartIdxWhere!=0
drhca7a26b2019-11-30 19:29:19 +00003379 && !whereUsablePartialIndex(pSrc->iCursor, isLeft, pWC,
3380 pProbe->pPartIdxWhere)
3381 ){
dan08291692014-08-27 17:37:20 +00003382 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */
drh4bd5f732013-07-31 23:22:39 +00003383 continue; /* Partial index inappropriate for this query */
3384 }
drh7e8515d2017-12-08 19:37:04 +00003385 if( pProbe->bNoQuery ) continue;
dan7de2a1f2014-04-28 20:11:20 +00003386 rSize = pProbe->aiRowLogEst[0];
drh5346e952013-05-08 14:14:26 +00003387 pNew->u.btree.nEq = 0;
dan71c57db2016-07-09 20:23:55 +00003388 pNew->u.btree.nBtm = 0;
3389 pNew->u.btree.nTop = 0;
drhc8bbce12014-10-21 01:05:09 +00003390 pNew->nSkip = 0;
drh4efc9292013-06-06 23:02:03 +00003391 pNew->nLTerm = 0;
drh23f98da2013-05-21 15:52:07 +00003392 pNew->iSortIdx = 0;
drhb8a8e8a2013-06-10 19:12:39 +00003393 pNew->rSetup = 0;
drh599d5762016-03-08 01:11:51 +00003394 pNew->prereq = mPrereq;
drh74f91d42013-06-19 18:01:44 +00003395 pNew->nOut = rSize;
drh23f98da2013-05-21 15:52:07 +00003396 pNew->u.btree.pIndex = pProbe;
3397 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor);
drh094afff2020-06-03 03:00:09 +00003398
drh53cfbe92013-06-13 17:28:22 +00003399 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */
3400 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 );
drh5f913ec2019-01-10 13:56:08 +00003401 if( pProbe->idxType==SQLITE_IDXTYPE_IPK ){
drh43fe25f2013-05-07 23:06:23 +00003402 /* Integer primary key index */
3403 pNew->wsFlags = WHERE_IPK;
drh23f98da2013-05-21 15:52:07 +00003404
3405 /* Full table scan */
drhd044d202013-05-31 12:43:55 +00003406 pNew->iSortIdx = b ? iSortIdx : 0;
drh40386962020-10-22 15:47:48 +00003407 /* TUNING: Cost of full table scan is 3.0*N. The 3.0 factor is an
3408 ** extra cost designed to discourage the use of full table scans,
3409 ** since index lookups have better worst-case performance if our
3410 ** stat guesses are wrong. Reduce the 3.0 penalty slightly
3411 ** (to 2.75) if we have valid STAT4 information for the table.
3412 ** At 2.75, a full table scan is preferred over using an index on
3413 ** a column with just two distinct values where each value has about
3414 ** an equal number of appearances. Without STAT4 data, we still want
3415 ** to use an index in that case, since the constraint might be for
3416 ** the scarcer of the two values, and in that case an index lookup is
3417 ** better.
3418 */
3419#ifdef SQLITE_ENABLE_STAT4
3420 pNew->rRun = rSize + 16 - 2*((pTab->tabFlags & TF_HasStat4)!=0);
3421#else
danaa9933c2014-04-24 20:04:49 +00003422 pNew->rRun = rSize + 16;
drh40386962020-10-22 15:47:48 +00003423#endif
drhdbd94862014-07-23 23:57:42 +00003424 ApplyCostMultiplier(pNew->rRun, pTab->costMult);
drhd8b77e22014-09-06 01:35:57 +00003425 whereLoopOutputAdjust(pWC, pNew, rSize);
drh23f98da2013-05-21 15:52:07 +00003426 rc = whereLoopInsert(pBuilder, pNew);
drhcca9f3d2013-09-06 15:23:29 +00003427 pNew->nOut = rSize;
drh23f98da2013-05-21 15:52:07 +00003428 if( rc ) break;
drh43fe25f2013-05-07 23:06:23 +00003429 }else{
drhec95c442013-10-23 01:57:32 +00003430 Bitmask m;
3431 if( pProbe->isCovering ){
3432 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED;
3433 m = 0;
3434 }else{
drh1fe3ac72018-06-09 01:12:08 +00003435 m = pSrc->colUsed & pProbe->colNotIdxed;
drhec95c442013-10-23 01:57:32 +00003436 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED;
3437 }
drh1c8148f2013-05-04 20:25:23 +00003438
drh23f98da2013-05-21 15:52:07 +00003439 /* Full scan via index */
drh53cfbe92013-06-13 17:28:22 +00003440 if( b
drh702ba9f2013-11-07 21:25:13 +00003441 || !HasRowid(pTab)
drh8dc570b2016-06-08 18:07:21 +00003442 || pProbe->pPartIdxWhere!=0
drh094afff2020-06-03 03:00:09 +00003443 || pSrc->fg.isIndexedBy
drh53cfbe92013-06-13 17:28:22 +00003444 || ( m==0
3445 && pProbe->bUnordered==0
drh702ba9f2013-11-07 21:25:13 +00003446 && (pProbe->szIdxRow<pTab->szTabRow)
drh53cfbe92013-06-13 17:28:22 +00003447 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0
3448 && sqlite3GlobalConfig.bUseCis
3449 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan)
3450 )
drhe3b7c922013-06-03 19:17:40 +00003451 ){
drh23f98da2013-05-21 15:52:07 +00003452 pNew->iSortIdx = b ? iSortIdx : 0;
danaa9933c2014-04-24 20:04:49 +00003453
3454 /* The cost of visiting the index rows is N*K, where K is
3455 ** between 1.1 and 3.0, depending on the relative sizes of the
drh2409f8a2016-07-27 18:27:02 +00003456 ** index and table rows. */
danaa9933c2014-04-24 20:04:49 +00003457 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow;
3458 if( m!=0 ){
drh2409f8a2016-07-27 18:27:02 +00003459 /* If this is a non-covering index scan, add in the cost of
3460 ** doing table lookups. The cost will be 3x the number of
3461 ** lookups. Take into account WHERE clause terms that can be
3462 ** satisfied using just the index, and that do not require a
3463 ** table lookup. */
3464 LogEst nLookup = rSize + 16; /* Base cost: N*3 */
3465 int ii;
3466 int iCur = pSrc->iCursor;
mistachkin19e76b22016-07-30 18:54:54 +00003467 WhereClause *pWC2 = &pWInfo->sWC;
3468 for(ii=0; ii<pWC2->nTerm; ii++){
3469 WhereTerm *pTerm = &pWC2->a[ii];
drh2409f8a2016-07-27 18:27:02 +00003470 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){
3471 break;
3472 }
3473 /* pTerm can be evaluated using just the index. So reduce
3474 ** the expected number of table lookups accordingly */
3475 if( pTerm->truthProb<=0 ){
3476 nLookup += pTerm->truthProb;
3477 }else{
3478 nLookup--;
3479 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19;
3480 }
3481 }
3482
3483 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup);
drhe1e2e9a2013-06-13 15:16:53 +00003484 }
drhdbd94862014-07-23 23:57:42 +00003485 ApplyCostMultiplier(pNew->rRun, pTab->costMult);
drhd8b77e22014-09-06 01:35:57 +00003486 whereLoopOutputAdjust(pWC, pNew, rSize);
drh23f98da2013-05-21 15:52:07 +00003487 rc = whereLoopInsert(pBuilder, pNew);
drhcca9f3d2013-09-06 15:23:29 +00003488 pNew->nOut = rSize;
drh23f98da2013-05-21 15:52:07 +00003489 if( rc ) break;
3490 }
3491 }
dan7a419232013-08-06 20:01:43 +00003492
drh89efac92020-02-22 16:58:49 +00003493 pBuilder->bldFlags1 = 0;
drhb8a8e8a2013-06-10 19:12:39 +00003494 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0);
drh89efac92020-02-22 16:58:49 +00003495 if( pBuilder->bldFlags1==SQLITE_BLDF1_INDEXED ){
drha3928dd2017-02-17 15:26:36 +00003496 /* If a non-unique index is used, or if a prefix of the key for
3497 ** unique index is used (making the index functionally non-unique)
3498 ** then the sqlite_stat1 data becomes important for scoring the
3499 ** plan */
3500 pTab->tabFlags |= TF_StatsUsed;
3501 }
drh175b8f02019-08-08 15:24:17 +00003502#ifdef SQLITE_ENABLE_STAT4
dan87cd9322013-08-07 15:52:41 +00003503 sqlite3Stat4ProbeFree(pBuilder->pRec);
3504 pBuilder->nRecValid = 0;
3505 pBuilder->pRec = 0;
danddc2d6e2013-08-06 20:15:06 +00003506#endif
drh1c8148f2013-05-04 20:25:23 +00003507 }
drh5346e952013-05-08 14:14:26 +00003508 return rc;
drhf1b5f5b2013-05-02 00:15:01 +00003509}
3510
drh8636e9c2013-06-11 01:50:08 +00003511#ifndef SQLITE_OMIT_VIRTUALTABLE
dan115305f2016-03-05 17:29:08 +00003512
3513/*
drh895bab32022-01-27 16:14:50 +00003514** Return true if pTerm is a virtual table LIMIT or OFFSET term.
3515*/
3516static int isLimitTerm(WhereTerm *pTerm){
drh8f2c0b52022-01-27 21:18:14 +00003517 assert( pTerm->eOperator==WO_AUX || pTerm->eMatchOp==0 );
3518 return pTerm->eMatchOp>=SQLITE_INDEX_CONSTRAINT_LIMIT
3519 && pTerm->eMatchOp<=SQLITE_INDEX_CONSTRAINT_OFFSET;
drh895bab32022-01-27 16:14:50 +00003520}
3521
3522/*
dan115305f2016-03-05 17:29:08 +00003523** Argument pIdxInfo is already populated with all constraints that may
3524** be used by the virtual table identified by pBuilder->pNew->iTab. This
3525** function marks a subset of those constraints usable, invokes the
3526** xBestIndex method and adds the returned plan to pBuilder.
3527**
3528** A constraint is marked usable if:
3529**
3530** * Argument mUsable indicates that its prerequisites are available, and
3531**
3532** * It is not one of the operators specified in the mExclude mask passed
3533** as the fourth argument (which in practice is either WO_IN or 0).
3534**
drh599d5762016-03-08 01:11:51 +00003535** Argument mPrereq is a mask of tables that must be scanned before the
dan115305f2016-03-05 17:29:08 +00003536** virtual table in question. These are added to the plans prerequisites
3537** before it is added to pBuilder.
3538**
3539** Output parameter *pbIn is set to true if the plan added to pBuilder
3540** uses one or more WO_IN terms, or false otherwise.
3541*/
3542static int whereLoopAddVirtualOne(
3543 WhereLoopBuilder *pBuilder,
drh8426e362016-03-08 01:32:30 +00003544 Bitmask mPrereq, /* Mask of tables that must be used. */
3545 Bitmask mUsable, /* Mask of usable tables */
3546 u16 mExclude, /* Exclude terms using these operators */
dan115305f2016-03-05 17:29:08 +00003547 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */
dan6256c1c2016-08-08 20:15:41 +00003548 u16 mNoOmit, /* Do not omit these constraints */
drh895bab32022-01-27 16:14:50 +00003549 int *pbIn, /* OUT: True if plan uses an IN(...) op */
3550 int *pbRetryLimit /* OUT: Retry without LIMIT/OFFSET */
dan115305f2016-03-05 17:29:08 +00003551){
3552 WhereClause *pWC = pBuilder->pWC;
drh0fe7e7d2022-02-01 14:58:29 +00003553 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1];
dan115305f2016-03-05 17:29:08 +00003554 struct sqlite3_index_constraint *pIdxCons;
3555 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage;
3556 int i;
3557 int mxTerm;
3558 int rc = SQLITE_OK;
3559 WhereLoop *pNew = pBuilder->pNew;
3560 Parse *pParse = pBuilder->pWInfo->pParse;
drh76012942021-02-21 21:04:54 +00003561 SrcItem *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab];
dan115305f2016-03-05 17:29:08 +00003562 int nConstraint = pIdxInfo->nConstraint;
3563
drh599d5762016-03-08 01:11:51 +00003564 assert( (mUsable & mPrereq)==mPrereq );
dan115305f2016-03-05 17:29:08 +00003565 *pbIn = 0;
drh599d5762016-03-08 01:11:51 +00003566 pNew->prereq = mPrereq;
dan115305f2016-03-05 17:29:08 +00003567
3568 /* Set the usable flag on the subset of constraints identified by
3569 ** arguments mUsable and mExclude. */
3570 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
3571 for(i=0; i<nConstraint; i++, pIdxCons++){
3572 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset];
3573 pIdxCons->usable = 0;
3574 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight
3575 && (pTerm->eOperator & mExclude)==0
drh895bab32022-01-27 16:14:50 +00003576 && (pbRetryLimit || !isLimitTerm(pTerm))
dan115305f2016-03-05 17:29:08 +00003577 ){
3578 pIdxCons->usable = 1;
3579 }
3580 }
3581
3582 /* Initialize the output fields of the sqlite3_index_info structure */
3583 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint);
drhd1cca3b2016-03-08 23:44:48 +00003584 assert( pIdxInfo->needToFreeIdxStr==0 );
dan115305f2016-03-05 17:29:08 +00003585 pIdxInfo->idxStr = 0;
3586 pIdxInfo->idxNum = 0;
dan115305f2016-03-05 17:29:08 +00003587 pIdxInfo->orderByConsumed = 0;
3588 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2;
3589 pIdxInfo->estimatedRows = 25;
3590 pIdxInfo->idxFlags = 0;
3591 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed;
drh0fe7e7d2022-02-01 14:58:29 +00003592 pHidden->mHandleIn = 0;
dan115305f2016-03-05 17:29:08 +00003593
3594 /* Invoke the virtual table xBestIndex() method */
3595 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo);
drh32dcc842018-11-16 13:56:15 +00003596 if( rc ){
3597 if( rc==SQLITE_CONSTRAINT ){
3598 /* If the xBestIndex method returns SQLITE_CONSTRAINT, that means
3599 ** that the particular combination of parameters provided is unusable.
3600 ** Make no entries in the loop table.
3601 */
drhe4f90b72018-11-16 15:08:31 +00003602 WHERETRACE(0xffff, (" ^^^^--- non-viable plan rejected!\n"));
drh32dcc842018-11-16 13:56:15 +00003603 return SQLITE_OK;
3604 }
3605 return rc;
3606 }
dan115305f2016-03-05 17:29:08 +00003607
3608 mxTerm = -1;
3609 assert( pNew->nLSlot>=nConstraint );
drh8f2c0b52022-01-27 21:18:14 +00003610 memset(pNew->aLTerm, 0, sizeof(pNew->aLTerm[0])*nConstraint );
3611 memset(&pNew->u.vtab, 0, sizeof(pNew->u.vtab));
dan115305f2016-03-05 17:29:08 +00003612 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
3613 for(i=0; i<nConstraint; i++, pIdxCons++){
3614 int iTerm;
3615 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){
3616 WhereTerm *pTerm;
3617 int j = pIdxCons->iTermOffset;
3618 if( iTerm>=nConstraint
3619 || j<0
3620 || j>=pWC->nTerm
3621 || pNew->aLTerm[iTerm]!=0
drh6de32e72016-03-08 02:59:33 +00003622 || pIdxCons->usable==0
dan115305f2016-03-05 17:29:08 +00003623 ){
drh6de32e72016-03-08 02:59:33 +00003624 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName);
drh337679b2018-04-09 15:57:54 +00003625 testcase( pIdxInfo->needToFreeIdxStr );
3626 return SQLITE_ERROR;
dan115305f2016-03-05 17:29:08 +00003627 }
3628 testcase( iTerm==nConstraint-1 );
3629 testcase( j==0 );
3630 testcase( j==pWC->nTerm-1 );
3631 pTerm = &pWC->a[j];
3632 pNew->prereq |= pTerm->prereqRight;
3633 assert( iTerm<pNew->nLSlot );
3634 pNew->aLTerm[iTerm] = pTerm;
3635 if( iTerm>mxTerm ) mxTerm = iTerm;
3636 testcase( iTerm==15 );
3637 testcase( iTerm==16 );
drh39593e42019-12-06 11:48:27 +00003638 if( pUsage[i].omit ){
3639 if( i<16 && ((1<<i)&mNoOmit)==0 ){
drhb6c94722019-12-05 21:46:23 +00003640 testcase( i!=iTerm );
3641 pNew->u.vtab.omitMask |= 1<<iTerm;
3642 }else{
3643 testcase( i!=iTerm );
3644 }
drh8f2c0b52022-01-27 21:18:14 +00003645 if( pTerm->eMatchOp==SQLITE_INDEX_CONSTRAINT_OFFSET ){
3646 pNew->u.vtab.bOmitOffset = 1;
3647 }
drhb6c94722019-12-05 21:46:23 +00003648 }
drh0fe7e7d2022-02-01 14:58:29 +00003649 if( SMASKBIT32(i) & pHidden->mHandleIn ){
drhb30298d2022-02-01 21:59:43 +00003650 pNew->u.vtab.mHandleIn |= MASKBIT32(iTerm);
drh0fe7e7d2022-02-01 14:58:29 +00003651 }else if( (pTerm->eOperator & WO_IN)!=0 ){
dan115305f2016-03-05 17:29:08 +00003652 /* A virtual table that is constrained by an IN clause may not
3653 ** consume the ORDER BY clause because (1) the order of IN terms
3654 ** is not necessarily related to the order of output terms and
3655 ** (2) Multiple outputs from a single IN value will not merge
3656 ** together. */
3657 pIdxInfo->orderByConsumed = 0;
3658 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE;
drh6de32e72016-03-08 02:59:33 +00003659 *pbIn = 1; assert( (mExclude & WO_IN)==0 );
dan115305f2016-03-05 17:29:08 +00003660 }
drh895bab32022-01-27 16:14:50 +00003661
3662 if( isLimitTerm(pTerm) && *pbIn ){
3663 /* If there is an IN(...) term handled as an == (separate call to
3664 ** xFilter for each value on the RHS of the IN) and a LIMIT or
3665 ** OFFSET term handled as well, the plan is unusable. Set output
3666 ** variable *pbRetryLimit to true to tell the caller to retry with
3667 ** LIMIT and OFFSET disabled. */
3668 if( pIdxInfo->needToFreeIdxStr ){
3669 sqlite3_free(pIdxInfo->idxStr);
3670 pIdxInfo->idxStr = 0;
3671 pIdxInfo->needToFreeIdxStr = 0;
3672 }
3673 *pbRetryLimit = 1;
3674 return SQLITE_OK;
3675 }
dan115305f2016-03-05 17:29:08 +00003676 }
3677 }
3678
3679 pNew->nLTerm = mxTerm+1;
drh337679b2018-04-09 15:57:54 +00003680 for(i=0; i<=mxTerm; i++){
3681 if( pNew->aLTerm[i]==0 ){
3682 /* The non-zero argvIdx values must be contiguous. Raise an
3683 ** error if they are not */
3684 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName);
3685 testcase( pIdxInfo->needToFreeIdxStr );
3686 return SQLITE_ERROR;
3687 }
3688 }
dan115305f2016-03-05 17:29:08 +00003689 assert( pNew->nLTerm<=pNew->nLSlot );
3690 pNew->u.vtab.idxNum = pIdxInfo->idxNum;
3691 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr;
3692 pIdxInfo->needToFreeIdxStr = 0;
3693 pNew->u.vtab.idxStr = pIdxInfo->idxStr;
3694 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ?
3695 pIdxInfo->nOrderBy : 0);
3696 pNew->rSetup = 0;
3697 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost);
3698 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows);
3699
3700 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated
3701 ** that the scan will visit at most one row. Clear it otherwise. */
3702 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){
3703 pNew->wsFlags |= WHERE_ONEROW;
3704 }else{
3705 pNew->wsFlags &= ~WHERE_ONEROW;
3706 }
drhbacbbcc2016-03-09 12:35:18 +00003707 rc = whereLoopInsert(pBuilder, pNew);
dan115305f2016-03-05 17:29:08 +00003708 if( pNew->u.vtab.needFree ){
3709 sqlite3_free(pNew->u.vtab.idxStr);
3710 pNew->u.vtab.needFree = 0;
3711 }
drh3349d9b2016-03-08 23:18:51 +00003712 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n",
3713 *pbIn, (sqlite3_uint64)mPrereq,
3714 (sqlite3_uint64)(pNew->prereq & ~mPrereq)));
dan115305f2016-03-05 17:29:08 +00003715
drhbacbbcc2016-03-09 12:35:18 +00003716 return rc;
dan115305f2016-03-05 17:29:08 +00003717}
3718
dane01b9282017-04-15 14:30:01 +00003719/*
drhb6592f62021-12-17 23:56:43 +00003720** Return the collating sequence for a constraint passed into xBestIndex.
3721**
3722** pIdxInfo must be an sqlite3_index_info structure passed into xBestIndex.
3723** This routine depends on there being a HiddenIndexInfo structure immediately
3724** following the sqlite3_index_info structure.
3725**
3726** Return a pointer to the collation name:
3727**
3728** 1. If there is an explicit COLLATE operator on the constaint, return it.
3729**
3730** 2. Else, if the column has an alternative collation, return that.
3731**
3732** 3. Otherwise, return "BINARY".
dane01b9282017-04-15 14:30:01 +00003733*/
drhefc88d02017-12-22 00:52:50 +00003734const char *sqlite3_vtab_collation(sqlite3_index_info *pIdxInfo, int iCons){
3735 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1];
dan0824ccf2017-04-14 19:41:37 +00003736 const char *zRet = 0;
drhefc88d02017-12-22 00:52:50 +00003737 if( iCons>=0 && iCons<pIdxInfo->nConstraint ){
dane42e1bc2017-12-19 18:56:28 +00003738 CollSeq *pC = 0;
drhefc88d02017-12-22 00:52:50 +00003739 int iTerm = pIdxInfo->aConstraint[iCons].iTermOffset;
3740 Expr *pX = pHidden->pWC->a[iTerm].pExpr;
dane42e1bc2017-12-19 18:56:28 +00003741 if( pX->pLeft ){
drh898c5272019-10-22 00:03:41 +00003742 pC = sqlite3ExprCompareCollSeq(pHidden->pParse, pX);
dane42e1bc2017-12-19 18:56:28 +00003743 }
drh7810ab62018-07-27 17:51:20 +00003744 zRet = (pC ? pC->zName : sqlite3StrBINARY);
dan0824ccf2017-04-14 19:41:37 +00003745 }
3746 return zRet;
3747}
3748
drhf1b5f5b2013-05-02 00:15:01 +00003749/*
drh0fe7e7d2022-02-01 14:58:29 +00003750** Return true if constraint iCons is really an IN(...) constraint, or
3751** false otherwise. If iCons is an IN(...) constraint, set (if bHandle!=0)
3752** or clear (if bHandle==0) the flag to handle it using an iterator.
3753*/
3754int sqlite3_vtab_in(sqlite3_index_info *pIdxInfo, int iCons, int bHandle){
3755 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1];
3756 u32 m = SMASKBIT32(iCons);
3757 if( m & pHidden->mIn ){
3758 if( bHandle==0 ){
3759 pHidden->mHandleIn &= ~m;
drhb30298d2022-02-01 21:59:43 +00003760 }else if( bHandle>0 ){
drh0fe7e7d2022-02-01 14:58:29 +00003761 pHidden->mHandleIn |= m;
3762 }
3763 return 1;
3764 }
3765 return 0;
3766}
3767
3768/*
drh82801a52022-01-20 17:10:59 +00003769** This interface is callable from within the xBestIndex callback only.
3770**
3771** If possible, set (*ppVal) to point to an object containing the value
3772** on the right-hand-side of constraint iCons.
3773*/
3774int sqlite3_vtab_rhs_value(
3775 sqlite3_index_info *pIdxInfo, /* Copy of first argument to xBestIndex */
3776 int iCons, /* Constraint for which RHS is wanted */
3777 sqlite3_value **ppVal /* Write value extracted here */
3778){
3779 HiddenIndexInfo *pH = (HiddenIndexInfo*)&pIdxInfo[1];
3780 sqlite3_value *pVal = 0;
3781 int rc = SQLITE_OK;
3782 if( iCons<0 || iCons>=pIdxInfo->nConstraint ){
drh991d1082022-01-21 00:38:49 +00003783 rc = SQLITE_MISUSE; /* EV: R-30545-25046 */
drh82801a52022-01-20 17:10:59 +00003784 }else{
3785 if( pH->aRhs[iCons]==0 ){
3786 WhereTerm *pTerm = &pH->pWC->a[pIdxInfo->aConstraint[iCons].iTermOffset];
3787 rc = sqlite3ValueFromExpr(
3788 pH->pParse->db, pTerm->pExpr->pRight, ENC(pH->pParse->db),
3789 SQLITE_AFF_BLOB, &pH->aRhs[iCons]
3790 );
drh991d1082022-01-21 00:38:49 +00003791 testcase( rc!=SQLITE_OK );
drh82801a52022-01-20 17:10:59 +00003792 }
3793 pVal = pH->aRhs[iCons];
3794 }
3795 *ppVal = pVal;
drh991d1082022-01-21 00:38:49 +00003796
drha1c81512022-01-21 18:57:30 +00003797 if( rc==SQLITE_OK && pVal==0 ){ /* IMP: R-19933-32160 */
drh991d1082022-01-21 00:38:49 +00003798 rc = SQLITE_NOTFOUND; /* IMP: R-36424-56542 */
3799 }
3800
drh82801a52022-01-20 17:10:59 +00003801 return rc;
3802}
3803
drhec778d22022-01-22 00:18:01 +00003804/*
3805** Return true if ORDER BY clause may be handled as DISTINCT.
3806*/
3807int sqlite3_vtab_distinct(sqlite3_index_info *pIdxInfo){
3808 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1];
drhd6df8552022-03-17 18:03:08 +00003809 assert( pHidden->eDistinct>=0 && pHidden->eDistinct<=3 );
drhec778d22022-01-22 00:18:01 +00003810 return pHidden->eDistinct;
3811}
3812
drh46dc6312022-03-09 14:22:28 +00003813#if (defined(SQLITE_ENABLE_DBPAGE_VTAB) || defined(SQLITE_TEST)) \
3814 && !defined(SQLITE_OMIT_VIRTUALTABLE)
3815/*
3816** Cause the prepared statement that is associated with a call to
drh7d0ae002022-04-08 17:01:29 +00003817** xBestIndex to potentiall use all schemas. If the statement being
3818** prepared is read-only, then just start read transactions on all
3819** schemas. But if this is a write operation, start writes on all
3820** schemas.
3821**
drh46dc6312022-03-09 14:22:28 +00003822** This is used by the (built-in) sqlite_dbpage virtual table.
3823*/
drh7d0ae002022-04-08 17:01:29 +00003824void sqlite3VtabUsesAllSchemas(sqlite3_index_info *pIdxInfo){
drh46dc6312022-03-09 14:22:28 +00003825 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1];
3826 Parse *pParse = pHidden->pParse;
drh6a51e702022-03-10 01:10:28 +00003827 int nDb = pParse->db->nDb;
3828 int i;
drh7d0ae002022-04-08 17:01:29 +00003829 for(i=0; i<nDb; i++){
3830 sqlite3CodeVerifySchema(pParse, i);
3831 }
3832 if( pParse->writeMask ){
3833 for(i=0; i<nDb; i++){
3834 sqlite3BeginWriteOperation(pParse, 0, i);
3835 }
3836 }
drh46dc6312022-03-09 14:22:28 +00003837}
3838#endif
3839
drh82801a52022-01-20 17:10:59 +00003840/*
drh0823c892013-05-11 00:06:23 +00003841** Add all WhereLoop objects for a table of the join identified by
3842** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table.
dan4f20cd42015-06-08 18:05:54 +00003843**
drh599d5762016-03-08 01:11:51 +00003844** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and
3845** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause
dan4f20cd42015-06-08 18:05:54 +00003846** entries that occur before the virtual table in the FROM clause and are
3847** separated from it by at least one LEFT or CROSS JOIN. Similarly, the
3848** mUnusable mask contains all FROM clause entries that occur after the
3849** virtual table and are separated from it by at least one LEFT or
3850** CROSS JOIN.
3851**
3852** For example, if the query were:
3853**
3854** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6;
3855**
drh599d5762016-03-08 01:11:51 +00003856** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6).
dan4f20cd42015-06-08 18:05:54 +00003857**
drh599d5762016-03-08 01:11:51 +00003858** All the tables in mPrereq must be scanned before the current virtual
dan4f20cd42015-06-08 18:05:54 +00003859** table. So any terms for which all prerequisites are satisfied by
drh599d5762016-03-08 01:11:51 +00003860** mPrereq may be specified as "usable" in all calls to xBestIndex.
dan4f20cd42015-06-08 18:05:54 +00003861** Conversely, all tables in mUnusable must be scanned after the current
3862** virtual table, so any terms for which the prerequisites overlap with
3863** mUnusable should always be configured as "not-usable" for xBestIndex.
drhf1b5f5b2013-05-02 00:15:01 +00003864*/
drh5346e952013-05-08 14:14:26 +00003865static int whereLoopAddVirtual(
danff4b23b2013-11-12 12:17:16 +00003866 WhereLoopBuilder *pBuilder, /* WHERE clause information */
drh599d5762016-03-08 01:11:51 +00003867 Bitmask mPrereq, /* Tables that must be scanned before this one */
dan4f20cd42015-06-08 18:05:54 +00003868 Bitmask mUnusable /* Tables that must be scanned after this one */
drhf1b5f5b2013-05-02 00:15:01 +00003869){
dan115305f2016-03-05 17:29:08 +00003870 int rc = SQLITE_OK; /* Return code */
drh70d18342013-06-06 19:16:33 +00003871 WhereInfo *pWInfo; /* WHERE analysis context */
drh5346e952013-05-08 14:14:26 +00003872 Parse *pParse; /* The parsing context */
3873 WhereClause *pWC; /* The WHERE clause */
drh76012942021-02-21 21:04:54 +00003874 SrcItem *pSrc; /* The FROM clause term to search */
dan115305f2016-03-05 17:29:08 +00003875 sqlite3_index_info *p; /* Object to pass to xBestIndex() */
3876 int nConstraint; /* Number of constraints in p */
3877 int bIn; /* True if plan uses IN(...) operator */
drh5346e952013-05-08 14:14:26 +00003878 WhereLoop *pNew;
dan115305f2016-03-05 17:29:08 +00003879 Bitmask mBest; /* Tables used by best possible plan */
dan6256c1c2016-08-08 20:15:41 +00003880 u16 mNoOmit;
drh895bab32022-01-27 16:14:50 +00003881 int bRetry = 0; /* True to retry with LIMIT/OFFSET disabled */
drh5346e952013-05-08 14:14:26 +00003882
drh599d5762016-03-08 01:11:51 +00003883 assert( (mPrereq & mUnusable)==0 );
drh70d18342013-06-06 19:16:33 +00003884 pWInfo = pBuilder->pWInfo;
3885 pParse = pWInfo->pParse;
drh5346e952013-05-08 14:14:26 +00003886 pWC = pBuilder->pWC;
drh5346e952013-05-08 14:14:26 +00003887 pNew = pBuilder->pNew;
drh70d18342013-06-06 19:16:33 +00003888 pSrc = &pWInfo->pTabList->a[pNew->iTab];
dan115305f2016-03-05 17:29:08 +00003889 assert( IsVirtual(pSrc->pTab) );
drhec778d22022-01-22 00:18:01 +00003890 p = allocateIndexInfo(pWInfo, pWC, mUnusable, pSrc, &mNoOmit);
dan115305f2016-03-05 17:29:08 +00003891 if( p==0 ) return SQLITE_NOMEM_BKPT;
drh5346e952013-05-08 14:14:26 +00003892 pNew->rSetup = 0;
3893 pNew->wsFlags = WHERE_VIRTUALTABLE;
drh4efc9292013-06-06 23:02:03 +00003894 pNew->nLTerm = 0;
drh5346e952013-05-08 14:14:26 +00003895 pNew->u.vtab.needFree = 0;
dan115305f2016-03-05 17:29:08 +00003896 nConstraint = p->nConstraint;
3897 if( whereLoopResize(pParse->db, pNew, nConstraint) ){
drh82801a52022-01-20 17:10:59 +00003898 freeIndexInfo(pParse->db, p);
mistachkinfad30392016-02-13 23:43:46 +00003899 return SQLITE_NOMEM_BKPT;
drh7963b0e2013-06-17 21:37:40 +00003900 }
drh5346e952013-05-08 14:14:26 +00003901
dan115305f2016-03-05 17:29:08 +00003902 /* First call xBestIndex() with all constraints usable. */
drh0f1631d2018-04-09 13:58:20 +00003903 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc->pTab->zName));
drh3349d9b2016-03-08 23:18:51 +00003904 WHERETRACE(0x40, (" VirtualOne: all usable\n"));
drh895bab32022-01-27 16:14:50 +00003905 rc = whereLoopAddVirtualOne(
3906 pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn, &bRetry
3907 );
3908 if( bRetry ){
3909 assert( rc==SQLITE_OK );
3910 rc = whereLoopAddVirtualOne(
3911 pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn, 0
3912 );
3913 }
dan076e0f92015-09-28 15:20:58 +00003914
dan115305f2016-03-05 17:29:08 +00003915 /* If the call to xBestIndex() with all terms enabled produced a plan
dan35808432019-03-29 13:17:50 +00003916 ** that does not require any source tables (IOW: a plan with mBest==0)
3917 ** and does not use an IN(...) operator, then there is no point in making
3918 ** any further calls to xBestIndex() since they will all return the same
3919 ** result (if the xBestIndex() implementation is sane). */
3920 if( rc==SQLITE_OK && ((mBest = (pNew->prereq & ~mPrereq))!=0 || bIn) ){
dan115305f2016-03-05 17:29:08 +00003921 int seenZero = 0; /* True if a plan with no prereqs seen */
3922 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */
3923 Bitmask mPrev = 0;
3924 Bitmask mBestNoIn = 0;
3925
3926 /* If the plan produced by the earlier call uses an IN(...) term, call
3927 ** xBestIndex again, this time with IN(...) terms disabled. */
drh3349d9b2016-03-08 23:18:51 +00003928 if( bIn ){
3929 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n"));
dan6256c1c2016-08-08 20:15:41 +00003930 rc = whereLoopAddVirtualOne(
drh895bab32022-01-27 16:14:50 +00003931 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn, 0);
drh6de32e72016-03-08 02:59:33 +00003932 assert( bIn==0 );
drh599d5762016-03-08 01:11:51 +00003933 mBestNoIn = pNew->prereq & ~mPrereq;
dan115305f2016-03-05 17:29:08 +00003934 if( mBestNoIn==0 ){
3935 seenZero = 1;
drh6de32e72016-03-08 02:59:33 +00003936 seenZeroNoIN = 1;
drh5346e952013-05-08 14:14:26 +00003937 }
3938 }
drh5346e952013-05-08 14:14:26 +00003939
drh599d5762016-03-08 01:11:51 +00003940 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq)
dan115305f2016-03-05 17:29:08 +00003941 ** in the set of terms that apply to the current virtual table. */
3942 while( rc==SQLITE_OK ){
3943 int i;
drh8426e362016-03-08 01:32:30 +00003944 Bitmask mNext = ALLBITS;
dan115305f2016-03-05 17:29:08 +00003945 assert( mNext>0 );
3946 for(i=0; i<nConstraint; i++){
3947 Bitmask mThis = (
drh599d5762016-03-08 01:11:51 +00003948 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq
dan115305f2016-03-05 17:29:08 +00003949 );
3950 if( mThis>mPrev && mThis<mNext ) mNext = mThis;
3951 }
3952 mPrev = mNext;
drh8426e362016-03-08 01:32:30 +00003953 if( mNext==ALLBITS ) break;
dan115305f2016-03-05 17:29:08 +00003954 if( mNext==mBest || mNext==mBestNoIn ) continue;
drh3349d9b2016-03-08 23:18:51 +00003955 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n",
3956 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext));
dan6256c1c2016-08-08 20:15:41 +00003957 rc = whereLoopAddVirtualOne(
drh895bab32022-01-27 16:14:50 +00003958 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn, 0);
drh599d5762016-03-08 01:11:51 +00003959 if( pNew->prereq==mPrereq ){
dan115305f2016-03-05 17:29:08 +00003960 seenZero = 1;
3961 if( bIn==0 ) seenZeroNoIN = 1;
3962 }
3963 }
3964
3965 /* If the calls to xBestIndex() in the above loop did not find a plan
3966 ** that requires no source tables at all (i.e. one guaranteed to be
3967 ** usable), make a call here with all source tables disabled */
3968 if( rc==SQLITE_OK && seenZero==0 ){
drh3349d9b2016-03-08 23:18:51 +00003969 WHERETRACE(0x40, (" VirtualOne: all disabled\n"));
dan6256c1c2016-08-08 20:15:41 +00003970 rc = whereLoopAddVirtualOne(
drh895bab32022-01-27 16:14:50 +00003971 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn, 0);
dan115305f2016-03-05 17:29:08 +00003972 if( bIn==0 ) seenZeroNoIN = 1;
3973 }
3974
3975 /* If the calls to xBestIndex() have so far failed to find a plan
3976 ** that requires no source tables at all and does not use an IN(...)
3977 ** operator, make a final call to obtain one here. */
3978 if( rc==SQLITE_OK && seenZeroNoIN==0 ){
drh3349d9b2016-03-08 23:18:51 +00003979 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n"));
dan6256c1c2016-08-08 20:15:41 +00003980 rc = whereLoopAddVirtualOne(
drh895bab32022-01-27 16:14:50 +00003981 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn, 0);
dan115305f2016-03-05 17:29:08 +00003982 }
3983 }
3984
3985 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr);
drh82801a52022-01-20 17:10:59 +00003986 freeIndexInfo(pParse->db, p);
drh0f1631d2018-04-09 13:58:20 +00003987 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc->pTab->zName, rc));
drh5346e952013-05-08 14:14:26 +00003988 return rc;
drhf1b5f5b2013-05-02 00:15:01 +00003989}
drh8636e9c2013-06-11 01:50:08 +00003990#endif /* SQLITE_OMIT_VIRTUALTABLE */
drhf1b5f5b2013-05-02 00:15:01 +00003991
3992/*
drhcf8fa7a2013-05-10 20:26:22 +00003993** Add WhereLoop entries to handle OR terms. This works for either
3994** btrees or virtual tables.
3995*/
dan4f20cd42015-06-08 18:05:54 +00003996static int whereLoopAddOr(
3997 WhereLoopBuilder *pBuilder,
drh599d5762016-03-08 01:11:51 +00003998 Bitmask mPrereq,
dan4f20cd42015-06-08 18:05:54 +00003999 Bitmask mUnusable
4000){
drh70d18342013-06-06 19:16:33 +00004001 WhereInfo *pWInfo = pBuilder->pWInfo;
drhcf8fa7a2013-05-10 20:26:22 +00004002 WhereClause *pWC;
4003 WhereLoop *pNew;
4004 WhereTerm *pTerm, *pWCEnd;
4005 int rc = SQLITE_OK;
4006 int iCur;
4007 WhereClause tempWC;
4008 WhereLoopBuilder sSubBuild;
dan5da73e12014-04-30 18:11:55 +00004009 WhereOrSet sSum, sCur;
drh76012942021-02-21 21:04:54 +00004010 SrcItem *pItem;
dan0824ccf2017-04-14 19:41:37 +00004011
drhcf8fa7a2013-05-10 20:26:22 +00004012 pWC = pBuilder->pWC;
drhcf8fa7a2013-05-10 20:26:22 +00004013 pWCEnd = pWC->a + pWC->nTerm;
4014 pNew = pBuilder->pNew;
drh77dfd5b2013-08-19 11:15:48 +00004015 memset(&sSum, 0, sizeof(sSum));
drh186ad8c2013-10-08 18:40:37 +00004016 pItem = pWInfo->pTabList->a + pNew->iTab;
4017 iCur = pItem->iCursor;
drhcf8fa7a2013-05-10 20:26:22 +00004018
drh61dac442022-04-12 14:23:45 +00004019 /* The multi-index OR optimization does not work for RIGHT and FULL JOIN */
4020 if( pItem->fg.jointype & JT_RIGHT ) return SQLITE_OK;
4021
drhcf8fa7a2013-05-10 20:26:22 +00004022 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){
4023 if( (pTerm->eOperator & WO_OR)!=0
4024 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0
4025 ){
4026 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc;
4027 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm];
4028 WhereTerm *pOrTerm;
drhaa32e3c2013-07-16 21:31:23 +00004029 int once = 1;
4030 int i, j;
dan0824ccf2017-04-14 19:41:37 +00004031
drh783dece2013-06-05 17:53:43 +00004032 sSubBuild = *pBuilder;
drhaa32e3c2013-07-16 21:31:23 +00004033 sSubBuild.pOrSet = &sCur;
drhcf8fa7a2013-05-10 20:26:22 +00004034
drh0a99ba32014-09-30 17:03:35 +00004035 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm));
drhc7f0d222013-06-19 03:27:12 +00004036 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){
drh783dece2013-06-05 17:53:43 +00004037 if( (pOrTerm->eOperator & WO_AND)!=0 ){
drhcf8fa7a2013-05-10 20:26:22 +00004038 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc;
4039 }else if( pOrTerm->leftCursor==iCur ){
drh70d18342013-06-06 19:16:33 +00004040 tempWC.pWInfo = pWC->pWInfo;
drh783dece2013-06-05 17:53:43 +00004041 tempWC.pOuter = pWC;
4042 tempWC.op = TK_AND;
drh783dece2013-06-05 17:53:43 +00004043 tempWC.nTerm = 1;
drh132f96f2021-12-08 16:07:22 +00004044 tempWC.nBase = 1;
drhcf8fa7a2013-05-10 20:26:22 +00004045 tempWC.a = pOrTerm;
4046 sSubBuild.pWC = &tempWC;
4047 }else{
4048 continue;
4049 }
drhaa32e3c2013-07-16 21:31:23 +00004050 sCur.n = 0;
drh52651492014-09-30 14:14:19 +00004051#ifdef WHERETRACE_ENABLED
drh0a99ba32014-09-30 17:03:35 +00004052 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n",
4053 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm));
4054 if( sqlite3WhereTrace & 0x400 ){
drhc84a4022016-05-27 12:30:20 +00004055 sqlite3WhereClausePrint(sSubBuild.pWC);
drh52651492014-09-30 14:14:19 +00004056 }
4057#endif
drh8636e9c2013-06-11 01:50:08 +00004058#ifndef SQLITE_OMIT_VIRTUALTABLE
drhcf8fa7a2013-05-10 20:26:22 +00004059 if( IsVirtual(pItem->pTab) ){
drh599d5762016-03-08 01:11:51 +00004060 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable);
drh8636e9c2013-06-11 01:50:08 +00004061 }else
4062#endif
4063 {
drh599d5762016-03-08 01:11:51 +00004064 rc = whereLoopAddBtree(&sSubBuild, mPrereq);
drhcf8fa7a2013-05-10 20:26:22 +00004065 }
drh36be4c42014-09-30 17:31:23 +00004066 if( rc==SQLITE_OK ){
drh599d5762016-03-08 01:11:51 +00004067 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable);
drh36be4c42014-09-30 17:31:23 +00004068 }
drh19c16c82021-04-16 12:13:39 +00004069 assert( rc==SQLITE_OK || rc==SQLITE_DONE || sCur.n==0
4070 || rc==SQLITE_NOMEM );
4071 testcase( rc==SQLITE_NOMEM && sCur.n>0 );
drh9a1f2e42019-12-28 03:55:50 +00004072 testcase( rc==SQLITE_DONE );
drhaa32e3c2013-07-16 21:31:23 +00004073 if( sCur.n==0 ){
4074 sSum.n = 0;
4075 break;
4076 }else if( once ){
4077 whereOrMove(&sSum, &sCur);
4078 once = 0;
4079 }else{
dan5da73e12014-04-30 18:11:55 +00004080 WhereOrSet sPrev;
drhaa32e3c2013-07-16 21:31:23 +00004081 whereOrMove(&sPrev, &sSum);
4082 sSum.n = 0;
4083 for(i=0; i<sPrev.n; i++){
4084 for(j=0; j<sCur.n; j++){
4085 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq,
drhbf539c42013-10-05 18:16:02 +00004086 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun),
4087 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut));
drhaa32e3c2013-07-16 21:31:23 +00004088 }
4089 }
4090 }
drhcf8fa7a2013-05-10 20:26:22 +00004091 }
drhaa32e3c2013-07-16 21:31:23 +00004092 pNew->nLTerm = 1;
4093 pNew->aLTerm[0] = pTerm;
4094 pNew->wsFlags = WHERE_MULTI_OR;
4095 pNew->rSetup = 0;
4096 pNew->iSortIdx = 0;
4097 memset(&pNew->u, 0, sizeof(pNew->u));
4098 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){
dan5da73e12014-04-30 18:11:55 +00004099 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs
4100 ** of all sub-scans required by the OR-scan. However, due to rounding
4101 ** errors, it may be that the cost of the OR-scan is equal to its
4102 ** most expensive sub-scan. Add the smallest possible penalty
4103 ** (equivalent to multiplying the cost by 1.07) to ensure that
4104 ** this does not happen. Otherwise, for WHERE clauses such as the
4105 ** following where there is an index on "y":
4106 **
4107 ** WHERE likelihood(x=?, 0.99) OR y=?
4108 **
4109 ** the planner may elect to "OR" together a full-table scan and an
4110 ** index lookup. And other similarly odd results. */
4111 pNew->rRun = sSum.a[i].rRun + 1;
drhaa32e3c2013-07-16 21:31:23 +00004112 pNew->nOut = sSum.a[i].nOut;
4113 pNew->prereq = sSum.a[i].prereq;
drhfd5874d2013-06-12 14:52:39 +00004114 rc = whereLoopInsert(pBuilder, pNew);
4115 }
drh0a99ba32014-09-30 17:03:35 +00004116 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm));
drhcf8fa7a2013-05-10 20:26:22 +00004117 }
4118 }
4119 return rc;
4120}
4121
4122/*
drhf1b5f5b2013-05-02 00:15:01 +00004123** Add all WhereLoop objects for all tables
4124*/
drh5346e952013-05-08 14:14:26 +00004125static int whereLoopAddAll(WhereLoopBuilder *pBuilder){
drh70d18342013-06-06 19:16:33 +00004126 WhereInfo *pWInfo = pBuilder->pWInfo;
drh599d5762016-03-08 01:11:51 +00004127 Bitmask mPrereq = 0;
drhf1b5f5b2013-05-02 00:15:01 +00004128 Bitmask mPrior = 0;
4129 int iTab;
drh70d18342013-06-06 19:16:33 +00004130 SrcList *pTabList = pWInfo->pTabList;
drh76012942021-02-21 21:04:54 +00004131 SrcItem *pItem;
4132 SrcItem *pEnd = &pTabList->a[pWInfo->nLevel];
drh70d18342013-06-06 19:16:33 +00004133 sqlite3 *db = pWInfo->pParse->db;
drh5346e952013-05-08 14:14:26 +00004134 int rc = SQLITE_OK;
drhb8a8e8a2013-06-10 19:12:39 +00004135 WhereLoop *pNew;
drhf1b5f5b2013-05-02 00:15:01 +00004136
4137 /* Loop over the tables in the join, from left to right */
drhb8a8e8a2013-06-10 19:12:39 +00004138 pNew = pBuilder->pNew;
drha2014152013-06-07 00:29:23 +00004139 whereLoopInit(pNew);
drh6fb5d352018-09-24 12:37:01 +00004140 pBuilder->iPlanLimit = SQLITE_QUERY_PLANNER_LIMIT;
dan4f20cd42015-06-08 18:05:54 +00004141 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){
4142 Bitmask mUnusable = 0;
drhb2a90f02013-05-10 03:30:49 +00004143 pNew->iTab = iTab;
drh6fb5d352018-09-24 12:37:01 +00004144 pBuilder->iPlanLimit += SQLITE_QUERY_PLANNER_LIMIT_INCR;
drh6f82e852015-06-06 20:12:09 +00004145 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor);
drha76ac882022-04-08 19:20:12 +00004146 if( (pItem->fg.jointype & (JT_OUTER|JT_CROSS))!=0 ){
dan4f20cd42015-06-08 18:05:54 +00004147 /* This condition is true when pItem is the FROM clause term on the
drha76ac882022-04-08 19:20:12 +00004148 ** right-hand-side of a OUTER or CROSS JOIN. */
drh599d5762016-03-08 01:11:51 +00004149 mPrereq = mPrior;
drhb1c993f2020-07-16 18:55:58 +00004150 }else{
4151 mPrereq = 0;
drhf1b5f5b2013-05-02 00:15:01 +00004152 }
drhec593592016-06-23 12:35:04 +00004153#ifndef SQLITE_OMIT_VIRTUALTABLE
drhb2a90f02013-05-10 03:30:49 +00004154 if( IsVirtual(pItem->pTab) ){
drh76012942021-02-21 21:04:54 +00004155 SrcItem *p;
dan4f20cd42015-06-08 18:05:54 +00004156 for(p=&pItem[1]; p<pEnd; p++){
drha76ac882022-04-08 19:20:12 +00004157 if( mUnusable || (p->fg.jointype & (JT_OUTER|JT_CROSS)) ){
dan4f20cd42015-06-08 18:05:54 +00004158 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor);
4159 }
4160 }
drh599d5762016-03-08 01:11:51 +00004161 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable);
drhec593592016-06-23 12:35:04 +00004162 }else
4163#endif /* SQLITE_OMIT_VIRTUALTABLE */
4164 {
drh599d5762016-03-08 01:11:51 +00004165 rc = whereLoopAddBtree(pBuilder, mPrereq);
drhb2a90f02013-05-10 03:30:49 +00004166 }
drhda230bd2018-06-09 00:09:58 +00004167 if( rc==SQLITE_OK && pBuilder->pWC->hasOr ){
drh599d5762016-03-08 01:11:51 +00004168 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable);
drhb2a90f02013-05-10 03:30:49 +00004169 }
drhb2a90f02013-05-10 03:30:49 +00004170 mPrior |= pNew->maskSelf;
drhfc9098a2018-09-21 18:43:51 +00004171 if( rc || db->mallocFailed ){
4172 if( rc==SQLITE_DONE ){
4173 /* We hit the query planner search limit set by iPlanLimit */
drh7ebb6052018-09-24 10:47:33 +00004174 sqlite3_log(SQLITE_WARNING, "abbreviated query algorithm search");
drhfc9098a2018-09-21 18:43:51 +00004175 rc = SQLITE_OK;
4176 }else{
4177 break;
4178 }
4179 }
drhf1b5f5b2013-05-02 00:15:01 +00004180 }
dan4f20cd42015-06-08 18:05:54 +00004181
drha2014152013-06-07 00:29:23 +00004182 whereLoopClear(db, pNew);
drh5346e952013-05-08 14:14:26 +00004183 return rc;
drhf1b5f5b2013-05-02 00:15:01 +00004184}
4185
drha18f3d22013-05-08 03:05:41 +00004186/*
drhc04ea802017-04-13 19:48:29 +00004187** Examine a WherePath (with the addition of the extra WhereLoop of the 6th
drh319f6772013-05-14 15:31:07 +00004188** parameters) to see if it outputs rows in the requested ORDER BY
drh0401ace2014-03-18 15:30:27 +00004189** (or GROUP BY) without requiring a separate sort operation. Return N:
drh319f6772013-05-14 15:31:07 +00004190**
drh0401ace2014-03-18 15:30:27 +00004191** N>0: N terms of the ORDER BY clause are satisfied
4192** N==0: No terms of the ORDER BY clause are satisfied
4193** N<0: Unknown yet how many terms of ORDER BY might be satisfied.
drh319f6772013-05-14 15:31:07 +00004194**
drh94433422013-07-01 11:05:50 +00004195** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as
4196** strict. With GROUP BY and DISTINCT the only requirement is that
4197** equivalent rows appear immediately adjacent to one another. GROUP BY
dan374cd782014-04-21 13:21:56 +00004198** and DISTINCT do not require rows to appear in any particular order as long
peter.d.reid60ec9142014-09-06 16:39:46 +00004199** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT
drh94433422013-07-01 11:05:50 +00004200** the pOrderBy terms can be matched in any order. With ORDER BY, the
4201** pOrderBy terms must be matched in strict left-to-right order.
drh6b7157b2013-05-10 02:00:35 +00004202*/
drh0401ace2014-03-18 15:30:27 +00004203static i8 wherePathSatisfiesOrderBy(
drh6b7157b2013-05-10 02:00:35 +00004204 WhereInfo *pWInfo, /* The WHERE clause */
drh4f402f22013-06-11 18:59:38 +00004205 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */
drh6b7157b2013-05-10 02:00:35 +00004206 WherePath *pPath, /* The WherePath to check */
drhd711e522016-05-19 22:40:04 +00004207 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */
drh4f402f22013-06-11 18:59:38 +00004208 u16 nLoop, /* Number of entries in pPath->aLoop[] */
drh319f6772013-05-14 15:31:07 +00004209 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */
drh4f402f22013-06-11 18:59:38 +00004210 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */
drh6b7157b2013-05-10 02:00:35 +00004211){
drh88da6442013-05-27 17:59:37 +00004212 u8 revSet; /* True if rev is known */
4213 u8 rev; /* Composite sort order */
4214 u8 revIdx; /* Index sort order */
drhe353ee32013-06-04 23:40:53 +00004215 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */
4216 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */
4217 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */
drhd711e522016-05-19 22:40:04 +00004218 u16 eqOpMask; /* Allowed equality operators */
drh416846a2013-11-06 12:56:04 +00004219 u16 nKeyCol; /* Number of key columns in pIndex */
4220 u16 nColumn; /* Total number of ordered columns in the index */
drh7699d1c2013-06-04 12:42:29 +00004221 u16 nOrderBy; /* Number terms in the ORDER BY clause */
4222 int iLoop; /* Index of WhereLoop in pPath being processed */
4223 int i, j; /* Loop counters */
4224 int iCur; /* Cursor number for current WhereLoop */
4225 int iColumn; /* A column number within table iCur */
drhe8ae5832013-06-19 13:32:46 +00004226 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */
drh7699d1c2013-06-04 12:42:29 +00004227 WhereTerm *pTerm; /* A single term of the WHERE clause */
4228 Expr *pOBExpr; /* An expression from the ORDER BY clause */
4229 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */
4230 Index *pIndex; /* The index associated with pLoop */
4231 sqlite3 *db = pWInfo->pParse->db; /* Database connection */
4232 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */
4233 Bitmask obDone; /* Mask of all ORDER BY terms */
drhe353ee32013-06-04 23:40:53 +00004234 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */
drhb8916be2013-06-14 02:51:48 +00004235 Bitmask ready; /* Mask of inner loops */
drh319f6772013-05-14 15:31:07 +00004236
4237 /*
drh7699d1c2013-06-04 12:42:29 +00004238 ** We say the WhereLoop is "one-row" if it generates no more than one
4239 ** row of output. A WhereLoop is one-row if all of the following are true:
drh319f6772013-05-14 15:31:07 +00004240 ** (a) All index columns match with WHERE_COLUMN_EQ.
4241 ** (b) The index is unique
drh7699d1c2013-06-04 12:42:29 +00004242 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row.
4243 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags.
drh319f6772013-05-14 15:31:07 +00004244 **
drhe353ee32013-06-04 23:40:53 +00004245 ** We say the WhereLoop is "order-distinct" if the set of columns from
4246 ** that WhereLoop that are in the ORDER BY clause are different for every
4247 ** row of the WhereLoop. Every one-row WhereLoop is automatically
4248 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause
4249 ** is not order-distinct. To be order-distinct is not quite the same as being
4250 ** UNIQUE since a UNIQUE column or index can have multiple rows that
4251 ** are NULL and NULL values are equivalent for the purpose of order-distinct.
4252 ** To be order-distinct, the columns must be UNIQUE and NOT NULL.
4253 **
4254 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the
4255 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is
4256 ** automatically order-distinct.
drh319f6772013-05-14 15:31:07 +00004257 */
4258
4259 assert( pOrderBy!=0 );
drh7699d1c2013-06-04 12:42:29 +00004260 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0;
drh319f6772013-05-14 15:31:07 +00004261
drh319f6772013-05-14 15:31:07 +00004262 nOrderBy = pOrderBy->nExpr;
drh7963b0e2013-06-17 21:37:40 +00004263 testcase( nOrderBy==BMS-1 );
drhe353ee32013-06-04 23:40:53 +00004264 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */
4265 isOrderDistinct = 1;
drh7699d1c2013-06-04 12:42:29 +00004266 obDone = MASKBIT(nOrderBy)-1;
drhe353ee32013-06-04 23:40:53 +00004267 orderDistinctMask = 0;
drhb8916be2013-06-14 02:51:48 +00004268 ready = 0;
drhd711e522016-05-19 22:40:04 +00004269 eqOpMask = WO_EQ | WO_IS | WO_ISNULL;
drh413b94a2020-07-10 19:09:40 +00004270 if( wctrlFlags & (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MAX|WHERE_ORDERBY_MIN) ){
4271 eqOpMask |= WO_IN;
4272 }
drhe353ee32013-06-04 23:40:53 +00004273 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){
drhb8916be2013-06-14 02:51:48 +00004274 if( iLoop>0 ) ready |= pLoop->maskSelf;
drhd711e522016-05-19 22:40:04 +00004275 if( iLoop<nLoop ){
4276 pLoop = pPath->aLoop[iLoop];
4277 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue;
4278 }else{
4279 pLoop = pLast;
4280 }
drh9dfaf622014-04-25 14:42:17 +00004281 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){
danece092e2022-03-16 12:06:00 +00004282 if( pLoop->u.vtab.isOrdered
4283 && ((wctrlFlags&(WHERE_DISTINCTBY|WHERE_SORTBYGROUP))!=WHERE_DISTINCTBY)
4284 ){
drhff1032e2019-11-08 20:13:44 +00004285 obSat = obDone;
4286 }
drh9dfaf622014-04-25 14:42:17 +00004287 break;
dana79a0e72019-07-29 14:42:56 +00004288 }else if( wctrlFlags & WHERE_DISTINCTBY ){
4289 pLoop->u.btree.nDistinctCol = 0;
drh9dfaf622014-04-25 14:42:17 +00004290 }
drh319f6772013-05-14 15:31:07 +00004291 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor;
drhb8916be2013-06-14 02:51:48 +00004292
4293 /* Mark off any ORDER BY term X that is a column in the table of
4294 ** the current loop for which there is term in the WHERE
4295 ** clause of the form X IS NULL or X=? that reference only outer
4296 ** loops.
4297 */
4298 for(i=0; i<nOrderBy; i++){
4299 if( MASKBIT(i) & obSat ) continue;
drh0d950af2019-08-22 16:38:42 +00004300 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr);
drh235667a2020-11-08 20:44:30 +00004301 if( NEVER(pOBExpr==0) ) continue;
dan4fcb30b2021-03-09 16:06:25 +00004302 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue;
drhb8916be2013-06-14 02:51:48 +00004303 if( pOBExpr->iTable!=iCur ) continue;
drh6f82e852015-06-06 20:12:09 +00004304 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn,
drhd711e522016-05-19 22:40:04 +00004305 ~ready, eqOpMask, 0);
drhb8916be2013-06-14 02:51:48 +00004306 if( pTerm==0 ) continue;
drh57a8c612016-09-07 01:51:46 +00004307 if( pTerm->eOperator==WO_IN ){
4308 /* IN terms are only valid for sorting in the ORDER BY LIMIT
4309 ** optimization, and then only if they are actually used
4310 ** by the query plan */
drh6e4b1402020-07-14 01:51:53 +00004311 assert( wctrlFlags &
4312 (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX) );
drh57a8c612016-09-07 01:51:46 +00004313 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){}
4314 if( j>=pLoop->nLTerm ) continue;
4315 }
drhe8d0c612015-05-14 01:05:25 +00004316 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){
dan41aa4422020-02-12 11:57:35 +00004317 Parse *pParse = pWInfo->pParse;
4318 CollSeq *pColl1 = sqlite3ExprNNCollSeq(pParse, pOrderBy->a[i].pExpr);
4319 CollSeq *pColl2 = sqlite3ExprCompareCollSeq(pParse, pTerm->pExpr);
drh77c9b3c2020-02-13 11:46:47 +00004320 assert( pColl1 );
dan41aa4422020-02-12 11:57:35 +00004321 if( pColl2==0 || sqlite3StrICmp(pColl1->zName, pColl2->zName) ){
drh70efa842017-09-28 01:58:23 +00004322 continue;
4323 }
drhe0cc3c22015-05-13 17:54:08 +00004324 testcase( pTerm->pExpr->op==TK_IS );
drhb8916be2013-06-14 02:51:48 +00004325 }
4326 obSat |= MASKBIT(i);
4327 }
4328
drh7699d1c2013-06-04 12:42:29 +00004329 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){
4330 if( pLoop->wsFlags & WHERE_IPK ){
4331 pIndex = 0;
drhbbbdc832013-10-22 18:01:40 +00004332 nKeyCol = 0;
drh416846a2013-11-06 12:56:04 +00004333 nColumn = 1;
drh7699d1c2013-06-04 12:42:29 +00004334 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){
drh1b0f0262013-05-30 22:27:09 +00004335 return 0;
drh7699d1c2013-06-04 12:42:29 +00004336 }else{
drhbbbdc832013-10-22 18:01:40 +00004337 nKeyCol = pIndex->nKeyCol;
drh416846a2013-11-06 12:56:04 +00004338 nColumn = pIndex->nColumn;
4339 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) );
drh4b92f982015-09-29 17:20:14 +00004340 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID
4341 || !HasRowid(pIndex->pTable));
drh8ed8ddf2021-04-26 14:32:48 +00004342 /* All relevant terms of the index must also be non-NULL in order
4343 ** for isOrderDistinct to be true. So the isOrderDistint value
4344 ** computed here might be a false positive. Corrections will be
4345 ** made at tag-20210426-1 below */
drh2ad07d92019-07-30 14:22:10 +00004346 isOrderDistinct = IsUniqueIndex(pIndex)
4347 && (pLoop->wsFlags & WHERE_SKIPSCAN)==0;
drh1b0f0262013-05-30 22:27:09 +00004348 }
drh7699d1c2013-06-04 12:42:29 +00004349
drh7699d1c2013-06-04 12:42:29 +00004350 /* Loop through all columns of the index and deal with the ones
4351 ** that are not constrained by == or IN.
4352 */
4353 rev = revSet = 0;
drhe353ee32013-06-04 23:40:53 +00004354 distinctColumns = 0;
drh416846a2013-11-06 12:56:04 +00004355 for(j=0; j<nColumn; j++){
dand49fd4e2016-07-27 19:33:04 +00004356 u8 bOnce = 1; /* True to run the ORDER BY search loop */
drh7699d1c2013-06-04 12:42:29 +00004357
dand49fd4e2016-07-27 19:33:04 +00004358 assert( j>=pLoop->u.btree.nEq
4359 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip)
4360 );
4361 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){
4362 u16 eOp = pLoop->aLTerm[j]->eOperator;
4363
4364 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when
danf7c92e82019-08-21 14:54:50 +00004365 ** doing WHERE_ORDERBY_LIMIT processing). Except, IS and ISNULL
4366 ** terms imply that the index is not UNIQUE NOT NULL in which case
4367 ** the loop need to be marked as not order-distinct because it can
4368 ** have repeated NULL rows.
dand49fd4e2016-07-27 19:33:04 +00004369 **
4370 ** If the current term is a column of an ((?,?) IN (SELECT...))
4371 ** expression for which the SELECT returns more than one column,
4372 ** check that it is the only column used by this loop. Otherwise,
4373 ** if it is one of two or more, none of the columns can be
danf7c92e82019-08-21 14:54:50 +00004374 ** considered to match an ORDER BY term.
4375 */
dand49fd4e2016-07-27 19:33:04 +00004376 if( (eOp & eqOpMask)!=0 ){
danf7c92e82019-08-21 14:54:50 +00004377 if( eOp & (WO_ISNULL|WO_IS) ){
4378 testcase( eOp & WO_ISNULL );
4379 testcase( eOp & WO_IS );
dand49fd4e2016-07-27 19:33:04 +00004380 testcase( isOrderDistinct );
4381 isOrderDistinct = 0;
4382 }
4383 continue;
drh64bcb8c2016-08-26 03:42:57 +00004384 }else if( ALWAYS(eOp & WO_IN) ){
4385 /* ALWAYS() justification: eOp is an equality operator due to the
4386 ** j<pLoop->u.btree.nEq constraint above. Any equality other
4387 ** than WO_IN is captured by the previous "if". So this one
4388 ** always has to be WO_IN. */
dand49fd4e2016-07-27 19:33:04 +00004389 Expr *pX = pLoop->aLTerm[j]->pExpr;
4390 for(i=j+1; i<pLoop->u.btree.nEq; i++){
4391 if( pLoop->aLTerm[i]->pExpr==pX ){
4392 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) );
4393 bOnce = 0;
4394 break;
4395 }
4396 }
drh7963b0e2013-06-17 21:37:40 +00004397 }
drh7699d1c2013-06-04 12:42:29 +00004398 }
4399
drhe353ee32013-06-04 23:40:53 +00004400 /* Get the column number in the table (iColumn) and sort order
4401 ** (revIdx) for the j-th column of the index.
drh7699d1c2013-06-04 12:42:29 +00004402 */
drh416846a2013-11-06 12:56:04 +00004403 if( pIndex ){
drh7699d1c2013-06-04 12:42:29 +00004404 iColumn = pIndex->aiColumn[j];
dan15750a22019-08-16 21:07:19 +00004405 revIdx = pIndex->aSortOrder[j] & KEYINFO_ORDER_DESC;
drh488e6192017-09-28 00:01:36 +00004406 if( iColumn==pIndex->pTable->iPKey ) iColumn = XN_ROWID;
drhdc3cd4b2013-05-30 23:21:20 +00004407 }else{
drh4b92f982015-09-29 17:20:14 +00004408 iColumn = XN_ROWID;
drh7699d1c2013-06-04 12:42:29 +00004409 revIdx = 0;
drhdc3cd4b2013-05-30 23:21:20 +00004410 }
drh7699d1c2013-06-04 12:42:29 +00004411
4412 /* An unconstrained column that might be NULL means that this
drh8ed8ddf2021-04-26 14:32:48 +00004413 ** WhereLoop is not well-ordered. tag-20210426-1
drh7699d1c2013-06-04 12:42:29 +00004414 */
drh8ed8ddf2021-04-26 14:32:48 +00004415 if( isOrderDistinct ){
4416 if( iColumn>=0
4417 && j>=pLoop->u.btree.nEq
4418 && pIndex->pTable->aCol[iColumn].notNull==0
4419 ){
4420 isOrderDistinct = 0;
4421 }
4422 if( iColumn==XN_EXPR ){
4423 isOrderDistinct = 0;
4424 }
4425 }
drh7699d1c2013-06-04 12:42:29 +00004426
4427 /* Find the ORDER BY term that corresponds to the j-th column
dan374cd782014-04-21 13:21:56 +00004428 ** of the index and mark that ORDER BY term off
drh7699d1c2013-06-04 12:42:29 +00004429 */
drhe353ee32013-06-04 23:40:53 +00004430 isMatch = 0;
drh7699d1c2013-06-04 12:42:29 +00004431 for(i=0; bOnce && i<nOrderBy; i++){
4432 if( MASKBIT(i) & obSat ) continue;
drh0d950af2019-08-22 16:38:42 +00004433 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr);
drh93ec45d2013-06-17 18:20:48 +00004434 testcase( wctrlFlags & WHERE_GROUPBY );
4435 testcase( wctrlFlags & WHERE_DISTINCTBY );
drh235667a2020-11-08 20:44:30 +00004436 if( NEVER(pOBExpr==0) ) continue;
drh4f402f22013-06-11 18:59:38 +00004437 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0;
drh488e6192017-09-28 00:01:36 +00004438 if( iColumn>=XN_ROWID ){
dan4fcb30b2021-03-09 16:06:25 +00004439 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue;
drhdae26fe2015-09-24 18:47:59 +00004440 if( pOBExpr->iTable!=iCur ) continue;
4441 if( pOBExpr->iColumn!=iColumn ) continue;
4442 }else{
drhdb8e68b2017-09-28 01:09:42 +00004443 Expr *pIdxExpr = pIndex->aColExpr->a[j].pExpr;
4444 if( sqlite3ExprCompareSkip(pOBExpr, pIdxExpr, iCur) ){
drhdae26fe2015-09-24 18:47:59 +00004445 continue;
4446 }
4447 }
dan62f6f512017-08-18 08:29:37 +00004448 if( iColumn!=XN_ROWID ){
drh70efa842017-09-28 01:58:23 +00004449 pColl = sqlite3ExprNNCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr);
drh7699d1c2013-06-04 12:42:29 +00004450 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue;
4451 }
dana79a0e72019-07-29 14:42:56 +00004452 if( wctrlFlags & WHERE_DISTINCTBY ){
4453 pLoop->u.btree.nDistinctCol = j+1;
4454 }
drhe353ee32013-06-04 23:40:53 +00004455 isMatch = 1;
drh7699d1c2013-06-04 12:42:29 +00004456 break;
4457 }
drh49290472014-10-11 02:12:58 +00004458 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){
drh59b8f2e2014-03-22 00:27:14 +00004459 /* Make sure the sort order is compatible in an ORDER BY clause.
4460 ** Sort order is irrelevant for a GROUP BY clause. */
4461 if( revSet ){
dan15750a22019-08-16 21:07:19 +00004462 if( (rev ^ revIdx)!=(pOrderBy->a[i].sortFlags&KEYINFO_ORDER_DESC) ){
4463 isMatch = 0;
4464 }
drh59b8f2e2014-03-22 00:27:14 +00004465 }else{
dan15750a22019-08-16 21:07:19 +00004466 rev = revIdx ^ (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC);
drh59b8f2e2014-03-22 00:27:14 +00004467 if( rev ) *pRevMask |= MASKBIT(iLoop);
4468 revSet = 1;
4469 }
4470 }
dan15750a22019-08-16 21:07:19 +00004471 if( isMatch && (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL) ){
4472 if( j==pLoop->u.btree.nEq ){
4473 pLoop->wsFlags |= WHERE_BIGNULL_SORT;
4474 }else{
4475 isMatch = 0;
4476 }
4477 }
drhe353ee32013-06-04 23:40:53 +00004478 if( isMatch ){
dan90b2fe62016-10-10 14:34:00 +00004479 if( iColumn==XN_ROWID ){
drh7963b0e2013-06-17 21:37:40 +00004480 testcase( distinctColumns==0 );
4481 distinctColumns = 1;
4482 }
drh7699d1c2013-06-04 12:42:29 +00004483 obSat |= MASKBIT(i);
drh7699d1c2013-06-04 12:42:29 +00004484 }else{
4485 /* No match found */
drhbbbdc832013-10-22 18:01:40 +00004486 if( j==0 || j<nKeyCol ){
drh7963b0e2013-06-17 21:37:40 +00004487 testcase( isOrderDistinct!=0 );
4488 isOrderDistinct = 0;
4489 }
drh7699d1c2013-06-04 12:42:29 +00004490 break;
4491 }
4492 } /* end Loop over all index columns */
drh81186b42013-06-18 01:52:41 +00004493 if( distinctColumns ){
4494 testcase( isOrderDistinct==0 );
4495 isOrderDistinct = 1;
4496 }
drh7699d1c2013-06-04 12:42:29 +00004497 } /* end-if not one-row */
4498
4499 /* Mark off any other ORDER BY terms that reference pLoop */
drhe353ee32013-06-04 23:40:53 +00004500 if( isOrderDistinct ){
4501 orderDistinctMask |= pLoop->maskSelf;
drh7699d1c2013-06-04 12:42:29 +00004502 for(i=0; i<nOrderBy; i++){
4503 Expr *p;
drh434a9312014-02-26 02:26:09 +00004504 Bitmask mTerm;
drh7699d1c2013-06-04 12:42:29 +00004505 if( MASKBIT(i) & obSat ) continue;
4506 p = pOrderBy->a[i].pExpr;
drh6c1f4ef2015-06-08 14:23:15 +00004507 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p);
drh434a9312014-02-26 02:26:09 +00004508 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue;
4509 if( (mTerm&~orderDistinctMask)==0 ){
drh7699d1c2013-06-04 12:42:29 +00004510 obSat |= MASKBIT(i);
4511 }
drh0afb4232013-05-31 13:36:32 +00004512 }
drh319f6772013-05-14 15:31:07 +00004513 }
drhb8916be2013-06-14 02:51:48 +00004514 } /* End the loop over all WhereLoops from outer-most down to inner-most */
drh36ed0342014-03-28 12:56:57 +00004515 if( obSat==obDone ) return (i8)nOrderBy;
drhd2de8612014-03-18 18:59:07 +00004516 if( !isOrderDistinct ){
4517 for(i=nOrderBy-1; i>0; i--){
drhc59ffa82021-10-04 15:08:49 +00004518 Bitmask m = ALWAYS(i<BMS) ? MASKBIT(i) - 1 : 0;
drhd2de8612014-03-18 18:59:07 +00004519 if( (obSat&m)==m ) return i;
4520 }
4521 return 0;
4522 }
drh319f6772013-05-14 15:31:07 +00004523 return -1;
drh6b7157b2013-05-10 02:00:35 +00004524}
4525
dan374cd782014-04-21 13:21:56 +00004526
4527/*
4528** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(),
4529** the planner assumes that the specified pOrderBy list is actually a GROUP
4530** BY clause - and so any order that groups rows as required satisfies the
4531** request.
4532**
4533** Normally, in this case it is not possible for the caller to determine
4534** whether or not the rows are really being delivered in sorted order, or
4535** just in some other order that provides the required grouping. However,
4536** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then
4537** this function may be called on the returned WhereInfo object. It returns
4538** true if the rows really will be sorted in the specified order, or false
4539** otherwise.
4540**
4541** For example, assuming:
4542**
4543** CREATE INDEX i1 ON t1(x, Y);
4544**
4545** then
4546**
4547** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1
4548** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0
4549*/
4550int sqlite3WhereIsSorted(WhereInfo *pWInfo){
danece092e2022-03-16 12:06:00 +00004551 assert( pWInfo->wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY) );
dan374cd782014-04-21 13:21:56 +00004552 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP );
4553 return pWInfo->sorted;
4554}
4555
drhd15cb172013-05-21 19:23:10 +00004556#ifdef WHERETRACE_ENABLED
4557/* For debugging use only: */
4558static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){
4559 static char zName[65];
4560 int i;
4561 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; }
4562 if( pLast ) zName[i++] = pLast->cId;
4563 zName[i] = 0;
4564 return zName;
4565}
4566#endif
4567
drh6b7157b2013-05-10 02:00:35 +00004568/*
dan50ae31e2014-08-08 16:52:28 +00004569** Return the cost of sorting nRow rows, assuming that the keys have
4570** nOrderby columns and that the first nSorted columns are already in
4571** order.
4572*/
4573static LogEst whereSortingCost(
4574 WhereInfo *pWInfo,
4575 LogEst nRow,
4576 int nOrderBy,
4577 int nSorted
4578){
4579 /* TUNING: Estimated cost of a full external sort, where N is
4580 ** the number of rows to sort is:
4581 **
4582 ** cost = (3.0 * N * log(N)).
4583 **
4584 ** Or, if the order-by clause has X terms but only the last Y
4585 ** terms are out of order, then block-sorting will reduce the
4586 ** sorting cost to:
4587 **
4588 ** cost = (3.0 * N * log(N)) * (Y/X)
4589 **
4590 ** The (Y/X) term is implemented using stack variable rScale
drh58d6f632020-08-24 23:44:27 +00004591 ** below.
4592 */
dan50ae31e2014-08-08 16:52:28 +00004593 LogEst rScale, rSortCost;
4594 assert( nOrderBy>0 && 66==sqlite3LogEst(100) );
4595 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66;
drhc3489bb2016-02-25 16:04:59 +00004596 rSortCost = nRow + rScale + 16;
dan50ae31e2014-08-08 16:52:28 +00004597
drhc3489bb2016-02-25 16:04:59 +00004598 /* Multiple by log(M) where M is the number of output rows.
drh58d6f632020-08-24 23:44:27 +00004599 ** Use the LIMIT for M if it is smaller. Or if this sort is for
drh75e6bbb2020-11-11 19:11:44 +00004600 ** a DISTINCT operator, M will be the number of distinct output
drh58d6f632020-08-24 23:44:27 +00004601 ** rows, so fudge it downwards a bit.
4602 */
drh8c098e62016-02-25 23:21:41 +00004603 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){
4604 nRow = pWInfo->iLimit;
drh58d6f632020-08-24 23:44:27 +00004605 }else if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT) ){
4606 /* TUNING: In the sort for a DISTINCT operator, assume that the DISTINCT
4607 ** reduces the number of output rows by a factor of 2 */
drh5f086dd2021-04-29 13:37:36 +00004608 if( nRow>10 ){ nRow -= 10; assert( 10==sqlite3LogEst(2) ); }
dan50ae31e2014-08-08 16:52:28 +00004609 }
drhc3489bb2016-02-25 16:04:59 +00004610 rSortCost += estLog(nRow);
dan50ae31e2014-08-08 16:52:28 +00004611 return rSortCost;
4612}
4613
4614/*
dan51576f42013-07-02 10:06:15 +00004615** Given the list of WhereLoop objects at pWInfo->pLoops, this routine
drha18f3d22013-05-08 03:05:41 +00004616** attempts to find the lowest cost path that visits each WhereLoop
4617** once. This path is then loaded into the pWInfo->a[].pWLoop fields.
4618**
drhc7f0d222013-06-19 03:27:12 +00004619** Assume that the total number of output rows that will need to be sorted
4620** will be nRowEst (in the 10*log2 representation). Or, ignore sorting
4621** costs if nRowEst==0.
4622**
drha18f3d22013-05-08 03:05:41 +00004623** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation
4624** error occurs.
4625*/
drhbf539c42013-10-05 18:16:02 +00004626static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){
drh783dece2013-06-05 17:53:43 +00004627 int mxChoice; /* Maximum number of simultaneous paths tracked */
drha18f3d22013-05-08 03:05:41 +00004628 int nLoop; /* Number of terms in the join */
drhe1e2e9a2013-06-13 15:16:53 +00004629 Parse *pParse; /* Parsing context */
drha18f3d22013-05-08 03:05:41 +00004630 sqlite3 *db; /* The database connection */
4631 int iLoop; /* Loop counter over the terms of the join */
4632 int ii, jj; /* Loop counters */
drhfde1e6b2013-09-06 17:45:42 +00004633 int mxI = 0; /* Index of next entry to replace */
drhd2de8612014-03-18 18:59:07 +00004634 int nOrderBy; /* Number of ORDER BY clause terms */
drhbf539c42013-10-05 18:16:02 +00004635 LogEst mxCost = 0; /* Maximum cost of a set of paths */
dan50ae31e2014-08-08 16:52:28 +00004636 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */
drha18f3d22013-05-08 03:05:41 +00004637 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */
4638 WherePath *aFrom; /* All nFrom paths at the previous level */
4639 WherePath *aTo; /* The nTo best paths at the current level */
4640 WherePath *pFrom; /* An element of aFrom[] that we are working on */
4641 WherePath *pTo; /* An element of aTo[] that we are working on */
4642 WhereLoop *pWLoop; /* One of the WhereLoop objects */
4643 WhereLoop **pX; /* Used to divy up the pSpace memory */
dan50ae31e2014-08-08 16:52:28 +00004644 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */
drha18f3d22013-05-08 03:05:41 +00004645 char *pSpace; /* Temporary memory used by this routine */
dane2c27852014-08-08 17:25:33 +00004646 int nSpace; /* Bytes of space allocated at pSpace */
drha18f3d22013-05-08 03:05:41 +00004647
drhe1e2e9a2013-06-13 15:16:53 +00004648 pParse = pWInfo->pParse;
4649 db = pParse->db;
drha18f3d22013-05-08 03:05:41 +00004650 nLoop = pWInfo->nLevel;
drhe1e2e9a2013-06-13 15:16:53 +00004651 /* TUNING: For simple queries, only the best path is tracked.
4652 ** For 2-way joins, the 5 best paths are followed.
4653 ** For joins of 3 or more tables, track the 10 best paths */
drh2504c6c2014-06-02 11:26:33 +00004654 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10);
drha18f3d22013-05-08 03:05:41 +00004655 assert( nLoop<=pWInfo->pTabList->nSrc );
drhddef5dc2014-08-07 16:50:00 +00004656 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst));
drha18f3d22013-05-08 03:05:41 +00004657
dan50ae31e2014-08-08 16:52:28 +00004658 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this
4659 ** case the purpose of this call is to estimate the number of rows returned
4660 ** by the overall query. Once this estimate has been obtained, the caller
4661 ** will invoke this function a second time, passing the estimate as the
4662 ** nRowEst parameter. */
4663 if( pWInfo->pOrderBy==0 || nRowEst==0 ){
4664 nOrderBy = 0;
4665 }else{
4666 nOrderBy = pWInfo->pOrderBy->nExpr;
4667 }
4668
4669 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */
dane2c27852014-08-08 17:25:33 +00004670 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2;
4671 nSpace += sizeof(LogEst) * nOrderBy;
drh575fad62016-02-05 13:38:36 +00004672 pSpace = sqlite3DbMallocRawNN(db, nSpace);
mistachkinfad30392016-02-13 23:43:46 +00004673 if( pSpace==0 ) return SQLITE_NOMEM_BKPT;
drha18f3d22013-05-08 03:05:41 +00004674 aTo = (WherePath*)pSpace;
4675 aFrom = aTo+mxChoice;
4676 memset(aFrom, 0, sizeof(aFrom[0]));
4677 pX = (WhereLoop**)(aFrom+mxChoice);
drhe9d935a2013-06-05 16:19:59 +00004678 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){
drha18f3d22013-05-08 03:05:41 +00004679 pFrom->aLoop = pX;
4680 }
dan50ae31e2014-08-08 16:52:28 +00004681 if( nOrderBy ){
4682 /* If there is an ORDER BY clause and it is not being ignored, set up
4683 ** space for the aSortCost[] array. Each element of the aSortCost array
4684 ** is either zero - meaning it has not yet been initialized - or the
4685 ** cost of sorting nRowEst rows of data where the first X terms of
4686 ** the ORDER BY clause are already in order, where X is the array
4687 ** index. */
4688 aSortCost = (LogEst*)pX;
dane2c27852014-08-08 17:25:33 +00004689 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy);
dan50ae31e2014-08-08 16:52:28 +00004690 }
dane2c27852014-08-08 17:25:33 +00004691 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] );
4692 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX );
drha18f3d22013-05-08 03:05:41 +00004693
drhe1e2e9a2013-06-13 15:16:53 +00004694 /* Seed the search with a single WherePath containing zero WhereLoops.
4695 **
danf104abb2015-03-16 20:40:00 +00004696 ** TUNING: Do not let the number of iterations go above 28. If the cost
4697 ** of computing an automatic index is not paid back within the first 28
drhe1e2e9a2013-06-13 15:16:53 +00004698 ** rows, then do not use the automatic index. */
danf104abb2015-03-16 20:40:00 +00004699 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) );
drha18f3d22013-05-08 03:05:41 +00004700 nFrom = 1;
dan50ae31e2014-08-08 16:52:28 +00004701 assert( aFrom[0].isOrdered==0 );
4702 if( nOrderBy ){
4703 /* If nLoop is zero, then there are no FROM terms in the query. Since
4704 ** in this case the query may return a maximum of one row, the results
4705 ** are already in the requested order. Set isOrdered to nOrderBy to
4706 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to
4707 ** -1, indicating that the result set may or may not be ordered,
4708 ** depending on the loops added to the current plan. */
4709 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy;
drh6b7157b2013-05-10 02:00:35 +00004710 }
4711
4712 /* Compute successively longer WherePaths using the previous generation
4713 ** of WherePaths as the basis for the next. Keep track of the mxChoice
4714 ** best paths at each generation */
drha18f3d22013-05-08 03:05:41 +00004715 for(iLoop=0; iLoop<nLoop; iLoop++){
4716 nTo = 0;
4717 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){
4718 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){
dan50ae31e2014-08-08 16:52:28 +00004719 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */
4720 LogEst rCost; /* Cost of path (pFrom+pWLoop) */
4721 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */
4722 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */
4723 Bitmask maskNew; /* Mask of src visited by (..) */
4724 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */
4725
drha18f3d22013-05-08 03:05:41 +00004726 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue;
4727 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue;
drh492ad132018-05-14 22:46:11 +00004728 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<3 ){
drh5a6f5ed2016-02-25 18:22:09 +00004729 /* Do not use an automatic index if the this loop is expected
drh492ad132018-05-14 22:46:11 +00004730 ** to run less than 1.25 times. It is tempting to also exclude
4731 ** automatic index usage on an outer loop, but sometimes an automatic
4732 ** index is useful in the outer loop of a correlated subquery. */
drh5a6f5ed2016-02-25 18:22:09 +00004733 assert( 10==sqlite3LogEst(2) );
drh87eb9192016-02-25 18:03:38 +00004734 continue;
4735 }
drh492ad132018-05-14 22:46:11 +00004736
drh6b7157b2013-05-10 02:00:35 +00004737 /* At this point, pWLoop is a candidate to be the next loop.
4738 ** Compute its cost */
dan50ae31e2014-08-08 16:52:28 +00004739 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow);
4740 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted);
drhfde1e6b2013-09-06 17:45:42 +00004741 nOut = pFrom->nRow + pWLoop->nOut;
drha18f3d22013-05-08 03:05:41 +00004742 maskNew = pFrom->maskLoop | pWLoop->maskSelf;
drh0401ace2014-03-18 15:30:27 +00004743 if( isOrdered<0 ){
4744 isOrdered = wherePathSatisfiesOrderBy(pWInfo,
drh4f402f22013-06-11 18:59:38 +00004745 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags,
drh0401ace2014-03-18 15:30:27 +00004746 iLoop, pWLoop, &revMask);
drh3a5ba8b2013-06-03 15:34:48 +00004747 }else{
4748 revMask = pFrom->revLoop;
drh6b7157b2013-05-10 02:00:35 +00004749 }
dan50ae31e2014-08-08 16:52:28 +00004750 if( isOrdered>=0 && isOrdered<nOrderBy ){
4751 if( aSortCost[isOrdered]==0 ){
4752 aSortCost[isOrdered] = whereSortingCost(
4753 pWInfo, nRowEst, nOrderBy, isOrdered
4754 );
4755 }
drhf559ed32018-07-28 21:01:55 +00004756 /* TUNING: Add a small extra penalty (5) to sorting as an
4757 ** extra encouragment to the query planner to select a plan
4758 ** where the rows emerge in the correct order without any sorting
4759 ** required. */
4760 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]) + 5;
dan50ae31e2014-08-08 16:52:28 +00004761
4762 WHERETRACE(0x002,
4763 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n",
4764 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy,
4765 rUnsorted, rCost));
4766 }else{
4767 rCost = rUnsorted;
drh54ac4452017-06-24 16:03:18 +00004768 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */
dan50ae31e2014-08-08 16:52:28 +00004769 }
4770
drhddef5dc2014-08-07 16:50:00 +00004771 /* Check to see if pWLoop should be added to the set of
4772 ** mxChoice best-so-far paths.
4773 **
4774 ** First look for an existing path among best-so-far paths
4775 ** that covers the same set of loops and has the same isOrdered
4776 ** setting as the current path candidate.
drhf2a90302014-08-07 20:37:01 +00004777 **
4778 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent
4779 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range
4780 ** of legal values for isOrdered, -1..64.
drhddef5dc2014-08-07 16:50:00 +00004781 */
drh6b7157b2013-05-10 02:00:35 +00004782 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){
drhfde1e6b2013-09-06 17:45:42 +00004783 if( pTo->maskLoop==maskNew
drhf2a90302014-08-07 20:37:01 +00004784 && ((pTo->isOrdered^isOrdered)&0x80)==0
drhfde1e6b2013-09-06 17:45:42 +00004785 ){
drh7963b0e2013-06-17 21:37:40 +00004786 testcase( jj==nTo-1 );
drh6b7157b2013-05-10 02:00:35 +00004787 break;
4788 }
4789 }
drha18f3d22013-05-08 03:05:41 +00004790 if( jj>=nTo ){
drhddef5dc2014-08-07 16:50:00 +00004791 /* None of the existing best-so-far paths match the candidate. */
drhddef5dc2014-08-07 16:50:00 +00004792 if( nTo>=mxChoice
dan50ae31e2014-08-08 16:52:28 +00004793 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted))
drhddef5dc2014-08-07 16:50:00 +00004794 ){
4795 /* The current candidate is no better than any of the mxChoice
4796 ** paths currently in the best-so-far buffer. So discard
4797 ** this candidate as not viable. */
drh989578e2013-10-28 14:34:35 +00004798#ifdef WHERETRACE_ENABLED /* 0x4 */
drhae70cf12013-05-31 15:18:46 +00004799 if( sqlite3WhereTrace&0x4 ){
drh78436d42017-05-22 00:45:15 +00004800 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n",
4801 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted,
drh0401ace2014-03-18 15:30:27 +00004802 isOrdered>=0 ? isOrdered+'0' : '?');
drhd15cb172013-05-21 19:23:10 +00004803 }
4804#endif
4805 continue;
4806 }
drhddef5dc2014-08-07 16:50:00 +00004807 /* If we reach this points it means that the new candidate path
4808 ** needs to be added to the set of best-so-far paths. */
drha18f3d22013-05-08 03:05:41 +00004809 if( nTo<mxChoice ){
drhd15cb172013-05-21 19:23:10 +00004810 /* Increase the size of the aTo set by one */
drha18f3d22013-05-08 03:05:41 +00004811 jj = nTo++;
4812 }else{
drhd15cb172013-05-21 19:23:10 +00004813 /* New path replaces the prior worst to keep count below mxChoice */
drhfde1e6b2013-09-06 17:45:42 +00004814 jj = mxI;
drha18f3d22013-05-08 03:05:41 +00004815 }
4816 pTo = &aTo[jj];
drh989578e2013-10-28 14:34:35 +00004817#ifdef WHERETRACE_ENABLED /* 0x4 */
drhae70cf12013-05-31 15:18:46 +00004818 if( sqlite3WhereTrace&0x4 ){
drh78436d42017-05-22 00:45:15 +00004819 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n",
4820 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted,
drh0401ace2014-03-18 15:30:27 +00004821 isOrdered>=0 ? isOrdered+'0' : '?');
drhd15cb172013-05-21 19:23:10 +00004822 }
4823#endif
drhf204dac2013-05-08 03:22:07 +00004824 }else{
drhddef5dc2014-08-07 16:50:00 +00004825 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the
drh78436d42017-05-22 00:45:15 +00004826 ** same set of loops and has the same isOrdered setting as the
drhddef5dc2014-08-07 16:50:00 +00004827 ** candidate path. Check to see if the candidate should replace
drh78436d42017-05-22 00:45:15 +00004828 ** pTo or if the candidate should be skipped.
4829 **
4830 ** The conditional is an expanded vector comparison equivalent to:
4831 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted)
4832 */
4833 if( pTo->rCost<rCost
4834 || (pTo->rCost==rCost
4835 && (pTo->nRow<nOut
4836 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted)
4837 )
4838 )
4839 ){
drh989578e2013-10-28 14:34:35 +00004840#ifdef WHERETRACE_ENABLED /* 0x4 */
drhae70cf12013-05-31 15:18:46 +00004841 if( sqlite3WhereTrace&0x4 ){
drhd15cb172013-05-21 19:23:10 +00004842 sqlite3DebugPrintf(
drh78436d42017-05-22 00:45:15 +00004843 "Skip %s cost=%-3d,%3d,%3d order=%c",
4844 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted,
drh0401ace2014-03-18 15:30:27 +00004845 isOrdered>=0 ? isOrdered+'0' : '?');
drh78436d42017-05-22 00:45:15 +00004846 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n",
drhfde1e6b2013-09-06 17:45:42 +00004847 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
drh78436d42017-05-22 00:45:15 +00004848 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?');
drhd15cb172013-05-21 19:23:10 +00004849 }
4850#endif
drhddef5dc2014-08-07 16:50:00 +00004851 /* Discard the candidate path from further consideration */
drh7963b0e2013-06-17 21:37:40 +00004852 testcase( pTo->rCost==rCost );
drhd15cb172013-05-21 19:23:10 +00004853 continue;
4854 }
drh7963b0e2013-06-17 21:37:40 +00004855 testcase( pTo->rCost==rCost+1 );
drhddef5dc2014-08-07 16:50:00 +00004856 /* Control reaches here if the candidate path is better than the
4857 ** pTo path. Replace pTo with the candidate. */
drh989578e2013-10-28 14:34:35 +00004858#ifdef WHERETRACE_ENABLED /* 0x4 */
drhae70cf12013-05-31 15:18:46 +00004859 if( sqlite3WhereTrace&0x4 ){
drhd15cb172013-05-21 19:23:10 +00004860 sqlite3DebugPrintf(
drh78436d42017-05-22 00:45:15 +00004861 "Update %s cost=%-3d,%3d,%3d order=%c",
4862 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted,
drh0401ace2014-03-18 15:30:27 +00004863 isOrdered>=0 ? isOrdered+'0' : '?');
drh78436d42017-05-22 00:45:15 +00004864 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n",
drhfde1e6b2013-09-06 17:45:42 +00004865 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
drh78436d42017-05-22 00:45:15 +00004866 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?');
drhd15cb172013-05-21 19:23:10 +00004867 }
4868#endif
drha18f3d22013-05-08 03:05:41 +00004869 }
drh6b7157b2013-05-10 02:00:35 +00004870 /* pWLoop is a winner. Add it to the set of best so far */
drha18f3d22013-05-08 03:05:41 +00004871 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf;
drh319f6772013-05-14 15:31:07 +00004872 pTo->revLoop = revMask;
drhfde1e6b2013-09-06 17:45:42 +00004873 pTo->nRow = nOut;
drha18f3d22013-05-08 03:05:41 +00004874 pTo->rCost = rCost;
dan50ae31e2014-08-08 16:52:28 +00004875 pTo->rUnsorted = rUnsorted;
drh6b7157b2013-05-10 02:00:35 +00004876 pTo->isOrdered = isOrdered;
drha18f3d22013-05-08 03:05:41 +00004877 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop);
4878 pTo->aLoop[iLoop] = pWLoop;
4879 if( nTo>=mxChoice ){
drhfde1e6b2013-09-06 17:45:42 +00004880 mxI = 0;
drha18f3d22013-05-08 03:05:41 +00004881 mxCost = aTo[0].rCost;
dan50ae31e2014-08-08 16:52:28 +00004882 mxUnsorted = aTo[0].nRow;
drha18f3d22013-05-08 03:05:41 +00004883 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){
dan50ae31e2014-08-08 16:52:28 +00004884 if( pTo->rCost>mxCost
4885 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted)
4886 ){
drhfde1e6b2013-09-06 17:45:42 +00004887 mxCost = pTo->rCost;
dan50ae31e2014-08-08 16:52:28 +00004888 mxUnsorted = pTo->rUnsorted;
drhfde1e6b2013-09-06 17:45:42 +00004889 mxI = jj;
4890 }
drha18f3d22013-05-08 03:05:41 +00004891 }
4892 }
4893 }
4894 }
4895
drh989578e2013-10-28 14:34:35 +00004896#ifdef WHERETRACE_ENABLED /* >=2 */
drh1b131b72014-10-21 16:01:40 +00004897 if( sqlite3WhereTrace & 0x02 ){
drha50ef112013-05-22 02:06:59 +00004898 sqlite3DebugPrintf("---- after round %d ----\n", iLoop);
drhd15cb172013-05-21 19:23:10 +00004899 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){
drhb8a8e8a2013-06-10 19:12:39 +00004900 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c",
drha50ef112013-05-22 02:06:59 +00004901 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
drh0401ace2014-03-18 15:30:27 +00004902 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?');
4903 if( pTo->isOrdered>0 ){
drh88da6442013-05-27 17:59:37 +00004904 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop);
4905 }else{
4906 sqlite3DebugPrintf("\n");
4907 }
drhf204dac2013-05-08 03:22:07 +00004908 }
4909 }
4910#endif
4911
drh6b7157b2013-05-10 02:00:35 +00004912 /* Swap the roles of aFrom and aTo for the next generation */
drha18f3d22013-05-08 03:05:41 +00004913 pFrom = aTo;
4914 aTo = aFrom;
4915 aFrom = pFrom;
4916 nFrom = nTo;
4917 }
4918
drh75b93402013-05-31 20:43:57 +00004919 if( nFrom==0 ){
drhe1e2e9a2013-06-13 15:16:53 +00004920 sqlite3ErrorMsg(pParse, "no query solution");
drhdbd6a7d2017-04-05 12:39:49 +00004921 sqlite3DbFreeNN(db, pSpace);
drh75b93402013-05-31 20:43:57 +00004922 return SQLITE_ERROR;
4923 }
drha18f3d22013-05-08 03:05:41 +00004924
drh6b7157b2013-05-10 02:00:35 +00004925 /* Find the lowest cost path. pFrom will be left pointing to that path */
drha18f3d22013-05-08 03:05:41 +00004926 pFrom = aFrom;
4927 for(ii=1; ii<nFrom; ii++){
4928 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii];
4929 }
4930 assert( pWInfo->nLevel==nLoop );
drh6b7157b2013-05-10 02:00:35 +00004931 /* Load the lowest cost path into pWInfo */
drha18f3d22013-05-08 03:05:41 +00004932 for(iLoop=0; iLoop<nLoop; iLoop++){
drh7ba39a92013-05-30 17:43:19 +00004933 WhereLevel *pLevel = pWInfo->a + iLoop;
4934 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop];
drhe217efc2013-06-12 03:48:41 +00004935 pLevel->iFrom = pWLoop->iTab;
drh7ba39a92013-05-30 17:43:19 +00004936 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor;
drha18f3d22013-05-08 03:05:41 +00004937 }
drhfd636c72013-06-21 02:05:06 +00004938 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0
4939 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0
4940 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP
drh4f402f22013-06-11 18:59:38 +00004941 && nRowEst
4942 ){
4943 Bitmask notUsed;
drh6457a352013-06-21 00:35:37 +00004944 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom,
drh93ec45d2013-06-17 18:20:48 +00004945 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], &notUsed);
drh0401ace2014-03-18 15:30:27 +00004946 if( rc==pWInfo->pResultSet->nExpr ){
4947 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED;
4948 }
drh4f402f22013-06-11 18:59:38 +00004949 }
drh6ee5a7b2018-09-08 20:09:46 +00004950 pWInfo->bOrderedInnerLoop = 0;
drh079a3072014-03-19 14:10:55 +00004951 if( pWInfo->pOrderBy ){
danece092e2022-03-16 12:06:00 +00004952 pWInfo->nOBSat = pFrom->isOrdered;
drh4f402f22013-06-11 18:59:38 +00004953 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){
drh079a3072014-03-19 14:10:55 +00004954 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){
4955 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED;
4956 }
drh4f402f22013-06-11 18:59:38 +00004957 }else{
drh4f402f22013-06-11 18:59:38 +00004958 pWInfo->revMask = pFrom->revLoop;
drha536df42016-05-19 22:13:37 +00004959 if( pWInfo->nOBSat<=0 ){
4960 pWInfo->nOBSat = 0;
drhc436a032016-10-12 18:55:53 +00004961 if( nLoop>0 ){
4962 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags;
4963 if( (wsFlags & WHERE_ONEROW)==0
4964 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN)
4965 ){
4966 Bitmask m = 0;
4967 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom,
drhd711e522016-05-19 22:40:04 +00004968 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m);
drhc436a032016-10-12 18:55:53 +00004969 testcase( wsFlags & WHERE_IPK );
4970 testcase( wsFlags & WHERE_COLUMN_IN );
4971 if( rc==pWInfo->pOrderBy->nExpr ){
4972 pWInfo->bOrderedInnerLoop = 1;
4973 pWInfo->revMask = m;
4974 }
drhd711e522016-05-19 22:40:04 +00004975 }
4976 }
drh751a44e2020-07-14 02:03:35 +00004977 }else if( nLoop
4978 && pWInfo->nOBSat==1
drh413b94a2020-07-10 19:09:40 +00004979 && (pWInfo->wctrlFlags & (WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX))!=0
4980 ){
drh19543b92020-07-14 22:20:26 +00004981 pWInfo->bOrderedInnerLoop = 1;
drha536df42016-05-19 22:13:37 +00004982 }
drh4f402f22013-06-11 18:59:38 +00004983 }
dan374cd782014-04-21 13:21:56 +00004984 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP)
drh11b04812015-04-12 01:22:04 +00004985 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0
dan374cd782014-04-21 13:21:56 +00004986 ){
danb6453202014-10-10 20:52:53 +00004987 Bitmask revMask = 0;
dan374cd782014-04-21 13:21:56 +00004988 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy,
danb6453202014-10-10 20:52:53 +00004989 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask
dan374cd782014-04-21 13:21:56 +00004990 );
4991 assert( pWInfo->sorted==0 );
danb6453202014-10-10 20:52:53 +00004992 if( nOrder==pWInfo->pOrderBy->nExpr ){
4993 pWInfo->sorted = 1;
4994 pWInfo->revMask = revMask;
4995 }
dan374cd782014-04-21 13:21:56 +00004996 }
drh6b7157b2013-05-10 02:00:35 +00004997 }
dan374cd782014-04-21 13:21:56 +00004998
4999
drha50ef112013-05-22 02:06:59 +00005000 pWInfo->nRowOut = pFrom->nRow;
drha18f3d22013-05-08 03:05:41 +00005001
5002 /* Free temporary memory and return success */
drhdbd6a7d2017-04-05 12:39:49 +00005003 sqlite3DbFreeNN(db, pSpace);
drha18f3d22013-05-08 03:05:41 +00005004 return SQLITE_OK;
5005}
drh94a11212004-09-25 13:12:14 +00005006
5007/*
drh60c96cd2013-06-09 17:21:25 +00005008** Most queries use only a single table (they are not joins) and have
5009** simple == constraints against indexed fields. This routine attempts
5010** to plan those simple cases using much less ceremony than the
5011** general-purpose query planner, and thereby yield faster sqlite3_prepare()
5012** times for the common case.
5013**
5014** Return non-zero on success, if this query can be handled by this
5015** no-frills query planner. Return zero if this query needs the
5016** general-purpose query planner.
5017*/
drhb8a8e8a2013-06-10 19:12:39 +00005018static int whereShortCut(WhereLoopBuilder *pBuilder){
drh60c96cd2013-06-09 17:21:25 +00005019 WhereInfo *pWInfo;
drh76012942021-02-21 21:04:54 +00005020 SrcItem *pItem;
drh60c96cd2013-06-09 17:21:25 +00005021 WhereClause *pWC;
5022 WhereTerm *pTerm;
5023 WhereLoop *pLoop;
5024 int iCur;
drh92a121f2013-06-10 12:15:47 +00005025 int j;
drh60c96cd2013-06-09 17:21:25 +00005026 Table *pTab;
5027 Index *pIdx;
drh36db90d2021-10-04 11:10:15 +00005028 WhereScan scan;
drh892ffcc2016-03-16 18:26:54 +00005029
drh60c96cd2013-06-09 17:21:25 +00005030 pWInfo = pBuilder->pWInfo;
drhce943bc2016-05-19 18:56:33 +00005031 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0;
drh60c96cd2013-06-09 17:21:25 +00005032 assert( pWInfo->pTabList->nSrc>=1 );
5033 pItem = pWInfo->pTabList->a;
5034 pTab = pItem->pTab;
5035 if( IsVirtual(pTab) ) return 0;
drh8a48b9c2015-08-19 15:20:00 +00005036 if( pItem->fg.isIndexedBy ) return 0;
drh60c96cd2013-06-09 17:21:25 +00005037 iCur = pItem->iCursor;
5038 pWC = &pWInfo->sWC;
5039 pLoop = pBuilder->pNew;
drh60c96cd2013-06-09 17:21:25 +00005040 pLoop->wsFlags = 0;
drhc8bbce12014-10-21 01:05:09 +00005041 pLoop->nSkip = 0;
drh36db90d2021-10-04 11:10:15 +00005042 pTerm = whereScanInit(&scan, pWC, iCur, -1, WO_EQ|WO_IS, 0);
drh3768f172021-10-06 10:04:04 +00005043 while( pTerm && pTerm->prereqRight ) pTerm = whereScanNext(&scan);
drh60c96cd2013-06-09 17:21:25 +00005044 if( pTerm ){
drhe8d0c612015-05-14 01:05:25 +00005045 testcase( pTerm->eOperator & WO_IS );
drh60c96cd2013-06-09 17:21:25 +00005046 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW;
5047 pLoop->aLTerm[0] = pTerm;
5048 pLoop->nLTerm = 1;
5049 pLoop->u.btree.nEq = 1;
drhe1e2e9a2013-06-13 15:16:53 +00005050 /* TUNING: Cost of a rowid lookup is 10 */
drhbf539c42013-10-05 18:16:02 +00005051 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */
drh60c96cd2013-06-09 17:21:25 +00005052 }else{
5053 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){
mistachkin4e5bef82015-05-15 20:14:00 +00005054 int opMask;
dancd40abb2013-08-29 10:46:05 +00005055 assert( pLoop->aLTermSpace==pLoop->aLTerm );
drh5f1d1d92014-07-31 22:59:04 +00005056 if( !IsUniqueIndex(pIdx)
dancd40abb2013-08-29 10:46:05 +00005057 || pIdx->pPartIdxWhere!=0
drhbbbdc832013-10-22 18:01:40 +00005058 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace)
dancd40abb2013-08-29 10:46:05 +00005059 ) continue;
mistachkin4e5bef82015-05-15 20:14:00 +00005060 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ;
drhbbbdc832013-10-22 18:01:40 +00005061 for(j=0; j<pIdx->nKeyCol; j++){
drh36db90d2021-10-04 11:10:15 +00005062 pTerm = whereScanInit(&scan, pWC, iCur, j, opMask, pIdx);
5063 while( pTerm && pTerm->prereqRight ) pTerm = whereScanNext(&scan);
drh60c96cd2013-06-09 17:21:25 +00005064 if( pTerm==0 ) break;
dan3072b532015-05-15 19:59:23 +00005065 testcase( pTerm->eOperator & WO_IS );
drh60c96cd2013-06-09 17:21:25 +00005066 pLoop->aLTerm[j] = pTerm;
5067 }
drhbbbdc832013-10-22 18:01:40 +00005068 if( j!=pIdx->nKeyCol ) continue;
drh92a121f2013-06-10 12:15:47 +00005069 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED;
drh1fe3ac72018-06-09 01:12:08 +00005070 if( pIdx->isCovering || (pItem->colUsed & pIdx->colNotIdxed)==0 ){
drh92a121f2013-06-10 12:15:47 +00005071 pLoop->wsFlags |= WHERE_IDX_ONLY;
5072 }
drh60c96cd2013-06-09 17:21:25 +00005073 pLoop->nLTerm = j;
5074 pLoop->u.btree.nEq = j;
5075 pLoop->u.btree.pIndex = pIdx;
drhe1e2e9a2013-06-13 15:16:53 +00005076 /* TUNING: Cost of a unique index lookup is 15 */
drhbf539c42013-10-05 18:16:02 +00005077 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */
drh60c96cd2013-06-09 17:21:25 +00005078 break;
5079 }
5080 }
drh3b75ffa2013-06-10 14:56:25 +00005081 if( pLoop->wsFlags ){
drhbf539c42013-10-05 18:16:02 +00005082 pLoop->nOut = (LogEst)1;
drh3b75ffa2013-06-10 14:56:25 +00005083 pWInfo->a[0].pWLoop = pLoop;
drh628dfe12017-04-03 14:07:08 +00005084 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] );
5085 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */
drh3b75ffa2013-06-10 14:56:25 +00005086 pWInfo->a[0].iTabCur = iCur;
5087 pWInfo->nRowOut = 1;
drhddba0c22014-03-18 20:33:42 +00005088 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr;
drh6457a352013-06-21 00:35:37 +00005089 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){
5090 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
5091 }
drh36db90d2021-10-04 11:10:15 +00005092 if( scan.iEquiv>1 ) pLoop->wsFlags |= WHERE_TRANSCONS;
drh3b75ffa2013-06-10 14:56:25 +00005093#ifdef SQLITE_DEBUG
5094 pLoop->cId = '0';
5095#endif
drh36db90d2021-10-04 11:10:15 +00005096#ifdef WHERETRACE_ENABLED
5097 if( sqlite3WhereTrace ){
5098 sqlite3DebugPrintf("whereShortCut() used to compute solution\n");
5099 }
5100#endif
drh3b75ffa2013-06-10 14:56:25 +00005101 return 1;
5102 }
5103 return 0;
drh60c96cd2013-06-09 17:21:25 +00005104}
5105
5106/*
danc456a762017-06-22 16:51:16 +00005107** Helper function for exprIsDeterministic().
5108*/
5109static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){
5110 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){
5111 pWalker->eCode = 0;
5112 return WRC_Abort;
5113 }
5114 return WRC_Continue;
5115}
5116
5117/*
5118** Return true if the expression contains no non-deterministic SQL
5119** functions. Do not consider non-deterministic SQL functions that are
5120** part of sub-select statements.
5121*/
5122static int exprIsDeterministic(Expr *p){
5123 Walker w;
5124 memset(&w, 0, sizeof(w));
5125 w.eCode = 1;
5126 w.xExprCallback = exprNodeIsDeterministic;
drh7e6f9802017-09-04 00:33:04 +00005127 w.xSelectCallback = sqlite3SelectWalkFail;
danc456a762017-06-22 16:51:16 +00005128 sqlite3WalkExpr(&w, p);
5129 return w.eCode;
5130}
5131
drhcea19512020-02-22 18:27:48 +00005132
5133#ifdef WHERETRACE_ENABLED
5134/*
5135** Display all WhereLoops in pWInfo
5136*/
5137static void showAllWhereLoops(WhereInfo *pWInfo, WhereClause *pWC){
5138 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */
5139 WhereLoop *p;
5140 int i;
5141 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz"
5142 "ABCDEFGHIJKLMNOPQRSTUVWYXZ";
5143 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){
5144 p->cId = zLabel[i%(sizeof(zLabel)-1)];
5145 sqlite3WhereLoopPrint(p, pWC);
5146 }
5147 }
5148}
5149# define WHERETRACE_ALL_LOOPS(W,C) showAllWhereLoops(W,C)
5150#else
5151# define WHERETRACE_ALL_LOOPS(W,C)
5152#endif
5153
drh70b403b2021-12-03 18:53:53 +00005154/* Attempt to omit tables from a join that do not affect the result.
5155** For a table to not affect the result, the following must be true:
5156**
5157** 1) The query must not be an aggregate.
5158** 2) The table must be the RHS of a LEFT JOIN.
5159** 3) Either the query must be DISTINCT, or else the ON or USING clause
5160** must contain a constraint that limits the scan of the table to
5161** at most a single row.
5162** 4) The table must not be referenced by any part of the query apart
5163** from its own USING or ON clause.
5164**
5165** For example, given:
5166**
5167** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1);
5168** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2);
5169** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3);
5170**
5171** then table t2 can be omitted from the following:
5172**
5173** SELECT v1, v3 FROM t1
5174** LEFT JOIN t2 ON (t1.ipk=t2.ipk)
5175** LEFT JOIN t3 ON (t1.ipk=t3.ipk)
5176**
5177** or from:
5178**
5179** SELECT DISTINCT v1, v3 FROM t1
5180** LEFT JOIN t2
5181** LEFT JOIN t3 ON (t1.ipk=t3.ipk)
5182*/
5183static SQLITE_NOINLINE Bitmask whereOmitNoopJoin(
5184 WhereInfo *pWInfo,
5185 Bitmask notReady
5186){
5187 int i;
5188 Bitmask tabUsed;
5189
5190 /* Preconditions checked by the caller */
5191 assert( pWInfo->nLevel>=2 );
5192 assert( OptimizationEnabled(pWInfo->pParse->db, SQLITE_OmitNoopJoin) );
5193
5194 /* These two preconditions checked by the caller combine to guarantee
5195 ** condition (1) of the header comment */
5196 assert( pWInfo->pResultSet!=0 );
5197 assert( 0==(pWInfo->wctrlFlags & WHERE_AGG_DISTINCT) );
5198
5199 tabUsed = sqlite3WhereExprListUsage(&pWInfo->sMaskSet, pWInfo->pResultSet);
5200 if( pWInfo->pOrderBy ){
5201 tabUsed |= sqlite3WhereExprListUsage(&pWInfo->sMaskSet, pWInfo->pOrderBy);
5202 }
5203 for(i=pWInfo->nLevel-1; i>=1; i--){
5204 WhereTerm *pTerm, *pEnd;
5205 SrcItem *pItem;
5206 WhereLoop *pLoop;
5207 pLoop = pWInfo->a[i].pWLoop;
5208 pItem = &pWInfo->pTabList->a[pLoop->iTab];
5209 if( (pItem->fg.jointype & JT_LEFT)==0 ) continue;
5210 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)==0
5211 && (pLoop->wsFlags & WHERE_ONEROW)==0
5212 ){
5213 continue;
5214 }
5215 if( (tabUsed & pLoop->maskSelf)!=0 ) continue;
5216 pEnd = pWInfo->sWC.a + pWInfo->sWC.nTerm;
5217 for(pTerm=pWInfo->sWC.a; pTerm<pEnd; pTerm++){
5218 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){
5219 if( !ExprHasProperty(pTerm->pExpr, EP_FromJoin)
drhd1985262022-04-11 11:25:28 +00005220 || pTerm->pExpr->w.iJoin!=pItem->iCursor
drh70b403b2021-12-03 18:53:53 +00005221 ){
5222 break;
5223 }
5224 }
5225 }
5226 if( pTerm<pEnd ) continue;
5227 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId));
5228 notReady &= ~pLoop->maskSelf;
5229 for(pTerm=pWInfo->sWC.a; pTerm<pEnd; pTerm++){
5230 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){
5231 pTerm->wtFlags |= TERM_CODED;
5232 }
5233 }
5234 if( i!=pWInfo->nLevel-1 ){
5235 int nByte = (pWInfo->nLevel-1-i) * sizeof(WhereLevel);
5236 memmove(&pWInfo->a[i], &pWInfo->a[i+1], nByte);
5237 }
5238 pWInfo->nLevel--;
5239 assert( pWInfo->nLevel>0 );
5240 }
5241 return notReady;
5242}
5243
danc456a762017-06-22 16:51:16 +00005244/*
drhfa35f5c2021-12-04 13:43:57 +00005245** Check to see if there are any SEARCH loops that might benefit from
5246** using a Bloom filter. Consider a Bloom filter if:
5247**
5248** (1) The SEARCH happens more than N times where N is the number
5249** of rows in the table that is being considered for the Bloom
drh7e910f62021-12-09 01:28:15 +00005250** filter.
5251** (2) Some searches are expected to find zero rows. (This is determined
5252** by the WHERE_SELFCULL flag on the term.)
drh5a4ac1c2021-12-09 19:42:52 +00005253** (3) Bloom-filter processing is not disabled. (Checked by the
drh7e910f62021-12-09 01:28:15 +00005254** caller.)
drh5a4ac1c2021-12-09 19:42:52 +00005255** (4) The size of the table being searched is known by ANALYZE.
drhfa35f5c2021-12-04 13:43:57 +00005256**
5257** This block of code merely checks to see if a Bloom filter would be
5258** appropriate, and if so sets the WHERE_BLOOMFILTER flag on the
5259** WhereLoop. The implementation of the Bloom filter comes further
5260** down where the code for each WhereLoop is generated.
5261*/
5262static SQLITE_NOINLINE void whereCheckIfBloomFilterIsUseful(
drhfecbf0a2021-12-04 21:11:18 +00005263 const WhereInfo *pWInfo
drhfa35f5c2021-12-04 13:43:57 +00005264){
5265 int i;
5266 LogEst nSearch;
drhfa35f5c2021-12-04 13:43:57 +00005267
5268 assert( pWInfo->nLevel>=2 );
drhfecbf0a2021-12-04 21:11:18 +00005269 assert( OptimizationEnabled(pWInfo->pParse->db, SQLITE_BloomFilter) );
drhfa35f5c2021-12-04 13:43:57 +00005270 nSearch = pWInfo->a[0].pWLoop->nOut;
5271 for(i=1; i<pWInfo->nLevel; i++){
5272 WhereLoop *pLoop = pWInfo->a[i].pWLoop;
drhb71a4852021-12-16 14:36:36 +00005273 const unsigned int reqFlags = (WHERE_SELFCULL|WHERE_COLUMN_EQ);
drhfb82caf2021-12-08 19:50:45 +00005274 if( (pLoop->wsFlags & reqFlags)==reqFlags
drh5a4ac1c2021-12-09 19:42:52 +00005275 /* vvvvvv--- Always the case if WHERE_COLUMN_EQ is defined */
drha11c5e22021-12-09 18:44:03 +00005276 && ALWAYS((pLoop->wsFlags & (WHERE_IPK|WHERE_INDEXED))!=0)
drhfa35f5c2021-12-04 13:43:57 +00005277 ){
drhfb82caf2021-12-08 19:50:45 +00005278 SrcItem *pItem = &pWInfo->pTabList->a[pLoop->iTab];
5279 Table *pTab = pItem->pTab;
5280 pTab->tabFlags |= TF_StatsUsed;
5281 if( nSearch > pTab->nRowLogEst
drh7e910f62021-12-09 01:28:15 +00005282 && (pTab->tabFlags & TF_HasStat1)!=0
drhfb82caf2021-12-08 19:50:45 +00005283 ){
drha11c5e22021-12-09 18:44:03 +00005284 testcase( pItem->fg.jointype & JT_LEFT );
drhfb82caf2021-12-08 19:50:45 +00005285 pLoop->wsFlags |= WHERE_BLOOMFILTER;
5286 pLoop->wsFlags &= ~WHERE_IDX_ONLY;
5287 WHERETRACE(0xffff, (
5288 "-> use Bloom-filter on loop %c because there are ~%.1e "
5289 "lookups into %s which has only ~%.1e rows\n",
5290 pLoop->cId, (double)sqlite3LogEstToInt(nSearch), pTab->zName,
5291 (double)sqlite3LogEstToInt(pTab->nRowLogEst)));
5292 }
drhfa35f5c2021-12-04 13:43:57 +00005293 }
5294 nSearch += pLoop->nOut;
5295 }
5296}
5297
5298/*
drhe3184742002-06-19 14:27:05 +00005299** Generate the beginning of the loop used for WHERE clause processing.
drhacf3b982005-01-03 01:27:18 +00005300** The return value is a pointer to an opaque structure that contains
drh75897232000-05-29 14:26:00 +00005301** information needed to terminate the loop. Later, the calling routine
danielk19774adee202004-05-08 08:23:19 +00005302** should invoke sqlite3WhereEnd() with the return value of this function
drh75897232000-05-29 14:26:00 +00005303** in order to complete the WHERE clause processing.
5304**
5305** If an error occurs, this routine returns NULL.
drhc27a1ce2002-06-14 20:58:45 +00005306**
5307** The basic idea is to do a nested loop, one loop for each table in
5308** the FROM clause of a select. (INSERT and UPDATE statements are the
5309** same as a SELECT with only a single table in the FROM clause.) For
5310** example, if the SQL is this:
5311**
5312** SELECT * FROM t1, t2, t3 WHERE ...;
5313**
5314** Then the code generated is conceptually like the following:
5315**
5316** foreach row1 in t1 do \ Code generated
danielk19774adee202004-05-08 08:23:19 +00005317** foreach row2 in t2 do |-- by sqlite3WhereBegin()
drhc27a1ce2002-06-14 20:58:45 +00005318** foreach row3 in t3 do /
5319** ...
5320** end \ Code generated
danielk19774adee202004-05-08 08:23:19 +00005321** end |-- by sqlite3WhereEnd()
drhc27a1ce2002-06-14 20:58:45 +00005322** end /
5323**
drh29dda4a2005-07-21 18:23:20 +00005324** Note that the loops might not be nested in the order in which they
5325** appear in the FROM clause if a different order is better able to make
drh51147ba2005-07-23 22:59:55 +00005326** use of indices. Note also that when the IN operator appears in
5327** the WHERE clause, it might result in additional nested loops for
5328** scanning through all values on the right-hand side of the IN.
drh29dda4a2005-07-21 18:23:20 +00005329**
drhc27a1ce2002-06-14 20:58:45 +00005330** There are Btree cursors associated with each table. t1 uses cursor
drh6a3ea0e2003-05-02 14:32:12 +00005331** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor.
5332** And so forth. This routine generates code to open those VDBE cursors
danielk19774adee202004-05-08 08:23:19 +00005333** and sqlite3WhereEnd() generates the code to close them.
drhc27a1ce2002-06-14 20:58:45 +00005334**
drhe6f85e72004-12-25 01:03:13 +00005335** The code that sqlite3WhereBegin() generates leaves the cursors named
5336** in pTabList pointing at their appropriate entries. The [...] code
drhf0863fe2005-06-12 21:35:51 +00005337** can use OP_Column and OP_Rowid opcodes on these cursors to extract
drhe6f85e72004-12-25 01:03:13 +00005338** data from the various tables of the loop.
5339**
drhc27a1ce2002-06-14 20:58:45 +00005340** If the WHERE clause is empty, the foreach loops must each scan their
5341** entire tables. Thus a three-way join is an O(N^3) operation. But if
5342** the tables have indices and there are terms in the WHERE clause that
5343** refer to those indices, a complete table scan can be avoided and the
5344** code will run much faster. Most of the work of this routine is checking
5345** to see if there are indices that can be used to speed up the loop.
5346**
5347** Terms of the WHERE clause are also used to limit which rows actually
5348** make it to the "..." in the middle of the loop. After each "foreach",
5349** terms of the WHERE clause that use only terms in that loop and outer
5350** loops are evaluated and if false a jump is made around all subsequent
5351** inner loops (or around the "..." if the test occurs within the inner-
5352** most loop)
5353**
5354** OUTER JOINS
5355**
5356** An outer join of tables t1 and t2 is conceptally coded as follows:
5357**
5358** foreach row1 in t1 do
5359** flag = 0
5360** foreach row2 in t2 do
5361** start:
5362** ...
5363** flag = 1
5364** end
drhe3184742002-06-19 14:27:05 +00005365** if flag==0 then
5366** move the row2 cursor to a null row
5367** goto start
5368** fi
drhc27a1ce2002-06-14 20:58:45 +00005369** end
5370**
drhe3184742002-06-19 14:27:05 +00005371** ORDER BY CLAUSE PROCESSING
5372**
drh94433422013-07-01 11:05:50 +00005373** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause
5374** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement
drhe3184742002-06-19 14:27:05 +00005375** if there is one. If there is no ORDER BY clause or if this routine
drh46ec5b62012-09-24 15:30:54 +00005376** is called from an UPDATE or DELETE statement, then pOrderBy is NULL.
drhfc8d4f92013-11-08 15:19:46 +00005377**
5378** The iIdxCur parameter is the cursor number of an index. If
drhce943bc2016-05-19 18:56:33 +00005379** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index
drhfc8d4f92013-11-08 15:19:46 +00005380** to use for OR clause processing. The WHERE clause should use this
5381** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is
5382** the first cursor in an array of cursors for all indices. iIdxCur should
5383** be used to compute the appropriate cursor depending on which index is
5384** used.
drh75897232000-05-29 14:26:00 +00005385*/
danielk19774adee202004-05-08 08:23:19 +00005386WhereInfo *sqlite3WhereBegin(
drhf1b5ff72016-04-14 13:35:26 +00005387 Parse *pParse, /* The parser context */
5388 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */
5389 Expr *pWhere, /* The WHERE clause */
5390 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */
drhe9ba9102017-02-16 20:52:52 +00005391 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */
drh895bab32022-01-27 16:14:50 +00005392 Select *pLimit, /* Use this LIMIT/OFFSET clause, if any */
drhf1b5ff72016-04-14 13:35:26 +00005393 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */
drhce943bc2016-05-19 18:56:33 +00005394 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number
drhf1b5ff72016-04-14 13:35:26 +00005395 ** If WHERE_USE_LIMIT, then the limit amount */
drh75897232000-05-29 14:26:00 +00005396){
danielk1977be229652009-03-20 14:18:51 +00005397 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */
drhc01a3c12009-12-16 22:10:49 +00005398 int nTabList; /* Number of elements in pTabList */
drh75897232000-05-29 14:26:00 +00005399 WhereInfo *pWInfo; /* Will become the return value of this function */
5400 Vdbe *v = pParse->pVdbe; /* The virtual database engine */
drhfe05af82005-07-21 03:14:59 +00005401 Bitmask notReady; /* Cursors that are not yet positioned */
drh1c8148f2013-05-04 20:25:23 +00005402 WhereLoopBuilder sWLB; /* The WhereLoop builder */
drh111a6a72008-12-21 03:51:16 +00005403 WhereMaskSet *pMaskSet; /* The expression mask set */
drh56f1b992012-09-25 14:29:39 +00005404 WhereLevel *pLevel; /* A single level in pWInfo->a[] */
drhfd636c72013-06-21 02:05:06 +00005405 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */
drh9cd1c992012-09-25 20:43:35 +00005406 int ii; /* Loop counter */
drh17435752007-08-16 04:30:38 +00005407 sqlite3 *db; /* Database connection */
drh5346e952013-05-08 14:14:26 +00005408 int rc; /* Return code */
drh9c0c57a2016-01-21 15:55:37 +00005409 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */
drh75897232000-05-29 14:26:00 +00005410
danf0ee1d32015-09-12 19:26:11 +00005411 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || (
5412 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0
drhce943bc2016-05-19 18:56:33 +00005413 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0
danf0ee1d32015-09-12 19:26:11 +00005414 ));
drh56f1b992012-09-25 14:29:39 +00005415
drhce943bc2016-05-19 18:56:33 +00005416 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */
5417 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0
drhc3489bb2016-02-25 16:04:59 +00005418 || (wctrlFlags & WHERE_USE_LIMIT)==0 );
5419
drh56f1b992012-09-25 14:29:39 +00005420 /* Variable initialization */
drhfd636c72013-06-21 02:05:06 +00005421 db = pParse->db;
drh1c8148f2013-05-04 20:25:23 +00005422 memset(&sWLB, 0, sizeof(sWLB));
drh0401ace2014-03-18 15:30:27 +00005423
5424 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */
5425 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 );
5426 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0;
drh56f1b992012-09-25 14:29:39 +00005427
drh29dda4a2005-07-21 18:23:20 +00005428 /* The number of tables in the FROM clause is limited by the number of
drh1398ad32005-01-19 23:24:50 +00005429 ** bits in a Bitmask
5430 */
drh67ae0cb2010-04-08 14:38:51 +00005431 testcase( pTabList->nSrc==BMS );
drh29dda4a2005-07-21 18:23:20 +00005432 if( pTabList->nSrc>BMS ){
5433 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS);
drh1398ad32005-01-19 23:24:50 +00005434 return 0;
5435 }
5436
drhc01a3c12009-12-16 22:10:49 +00005437 /* This function normally generates a nested loop for all tables in
drhce943bc2016-05-19 18:56:33 +00005438 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should
drhc01a3c12009-12-16 22:10:49 +00005439 ** only generate code for the first table in pTabList and assume that
5440 ** any cursors associated with subsequent tables are uninitialized.
5441 */
drhce943bc2016-05-19 18:56:33 +00005442 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc;
drhc01a3c12009-12-16 22:10:49 +00005443
drh75897232000-05-29 14:26:00 +00005444 /* Allocate and initialize the WhereInfo structure that will become the
danielk1977be229652009-03-20 14:18:51 +00005445 ** return value. A single allocation is used to store the WhereInfo
5446 ** struct, the contents of WhereInfo.a[], the WhereClause structure
5447 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte
5448 ** field (type Bitmask) it must be aligned on an 8-byte boundary on
5449 ** some architectures. Hence the ROUND8() below.
drh75897232000-05-29 14:26:00 +00005450 */
drhcf6e3fd2022-04-01 18:45:11 +00005451 nByteWInfo = ROUND8P(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel));
drh87c05f02016-10-03 14:44:47 +00005452 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop));
drh17435752007-08-16 04:30:38 +00005453 if( db->mallocFailed ){
drh8b307fb2010-04-06 15:57:05 +00005454 sqlite3DbFree(db, pWInfo);
5455 pWInfo = 0;
danielk197785574e32008-10-06 05:32:18 +00005456 goto whereBeginError;
drh75897232000-05-29 14:26:00 +00005457 }
5458 pWInfo->pParse = pParse;
5459 pWInfo->pTabList = pTabList;
drh6b7157b2013-05-10 02:00:35 +00005460 pWInfo->pOrderBy = pOrderBy;
drhaca19e12017-04-07 19:41:31 +00005461 pWInfo->pWhere = pWhere;
drh6457a352013-06-21 00:35:37 +00005462 pWInfo->pResultSet = pResultSet;
drh87c05f02016-10-03 14:44:47 +00005463 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1;
5464 pWInfo->nLevel = nTabList;
drhec4ccdb2018-12-29 02:26:59 +00005465 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(pParse);
drh6df2acd2008-12-28 16:55:25 +00005466 pWInfo->wctrlFlags = wctrlFlags;
drhc3489bb2016-02-25 16:04:59 +00005467 pWInfo->iLimit = iAuxArg;
drh8b307fb2010-04-06 15:57:05 +00005468 pWInfo->savedNQueryLoop = pParse->nQueryLoop;
drh895bab32022-01-27 16:14:50 +00005469#ifndef SQLITE_OMIT_VIRTUALTABLE
5470 pWInfo->pLimit = pLimit;
5471#endif
drh87c05f02016-10-03 14:44:47 +00005472 memset(&pWInfo->nOBSat, 0,
5473 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat));
5474 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel));
drhb0264ee2015-09-14 14:45:50 +00005475 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */
drh70d18342013-06-06 19:16:33 +00005476 pMaskSet = &pWInfo->sMaskSet;
drh844a89b2021-12-02 12:34:05 +00005477 pMaskSet->n = 0;
5478 pMaskSet->ix[0] = -99; /* Initialize ix[0] to a value that can never be
5479 ** a valid cursor number, to avoid an initial
5480 ** test for pMaskSet->n==0 in sqlite3WhereGetMask() */
drh1c8148f2013-05-04 20:25:23 +00005481 sWLB.pWInfo = pWInfo;
drh70d18342013-06-06 19:16:33 +00005482 sWLB.pWC = &pWInfo->sWC;
drh1ac87e12013-07-18 14:50:56 +00005483 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo);
5484 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) );
drh60c96cd2013-06-09 17:21:25 +00005485 whereLoopInit(sWLB.pNew);
drhb8a8e8a2013-06-10 19:12:39 +00005486#ifdef SQLITE_DEBUG
5487 sWLB.pNew->cId = '*';
5488#endif
drh08192d52002-04-30 19:20:28 +00005489
drh111a6a72008-12-21 03:51:16 +00005490 /* Split the WHERE clause into separate subexpressions where each
5491 ** subexpression is separated by an AND operator.
5492 */
drh6c1f4ef2015-06-08 14:23:15 +00005493 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo);
5494 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND);
drh111a6a72008-12-21 03:51:16 +00005495
drh4fe425a2013-06-12 17:08:06 +00005496 /* Special case: No FROM clause
5497 */
5498 if( nTabList==0 ){
drhddba0c22014-03-18 20:33:42 +00005499 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr;
drh9b3bfa02021-12-02 01:30:16 +00005500 if( (wctrlFlags & WHERE_WANT_DISTINCT)!=0
5501 && OptimizationEnabled(db, SQLITE_DistinctOpt)
5502 ){
drh6457a352013-06-21 00:35:37 +00005503 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
5504 }
drhfa16f5d2018-05-03 01:37:13 +00005505 ExplainQueryPlan((pParse, 0, "SCAN CONSTANT ROW"));
drh83e8ca52017-08-25 13:34:18 +00005506 }else{
5507 /* Assign a bit from the bitmask to every term in the FROM clause.
5508 **
5509 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N.
5510 **
5511 ** The rule of the previous sentence ensures thta if X is the bitmask for
5512 ** a table T, then X-1 is the bitmask for all other tables to the left of T.
5513 ** Knowing the bitmask for all tables to the left of a left join is
5514 ** important. Ticket #3015.
5515 **
5516 ** Note that bitmasks are created for all pTabList->nSrc tables in
5517 ** pTabList, not just the first nTabList tables. nTabList is normally
5518 ** equal to pTabList->nSrc but might be shortened to 1 if the
5519 ** WHERE_OR_SUBCLAUSE flag is set.
5520 */
5521 ii = 0;
5522 do{
5523 createMask(pMaskSet, pTabList->a[ii].iCursor);
5524 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC);
5525 }while( (++ii)<pTabList->nSrc );
5526 #ifdef SQLITE_DEBUG
5527 {
5528 Bitmask mx = 0;
5529 for(ii=0; ii<pTabList->nSrc; ii++){
5530 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor);
5531 assert( m>=mx );
5532 mx = m;
5533 }
drh269ba802017-07-04 19:34:36 +00005534 }
drh83e8ca52017-08-25 13:34:18 +00005535 #endif
drh4fe425a2013-06-12 17:08:06 +00005536 }
drh83e8ca52017-08-25 13:34:18 +00005537
drhb121dd12015-06-06 18:30:17 +00005538 /* Analyze all of the subexpressions. */
drh6c1f4ef2015-06-08 14:23:15 +00005539 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC);
drh895bab32022-01-27 16:14:50 +00005540 sqlite3WhereAddLimit(&pWInfo->sWC, pLimit);
drhb121dd12015-06-06 18:30:17 +00005541 if( db->mallocFailed ) goto whereBeginError;
drh75897232000-05-29 14:26:00 +00005542
danc456a762017-06-22 16:51:16 +00005543 /* Special case: WHERE terms that do not refer to any tables in the join
5544 ** (constant expressions). Evaluate each such term, and jump over all the
5545 ** generated code if the result is not true.
5546 **
5547 ** Do not do this if the expression contains non-deterministic functions
5548 ** that are not within a sub-select. This is not strictly required, but
5549 ** preserves SQLite's legacy behaviour in the following two cases:
5550 **
5551 ** FROM ... WHERE random()>0; -- eval random() once per row
5552 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall
5553 */
drh132f96f2021-12-08 16:07:22 +00005554 for(ii=0; ii<sWLB.pWC->nBase; ii++){
danc456a762017-06-22 16:51:16 +00005555 WhereTerm *pT = &sWLB.pWC->a[ii];
drh33f10202018-01-27 05:40:10 +00005556 if( pT->wtFlags & TERM_VIRTUAL ) continue;
danc456a762017-06-22 16:51:16 +00005557 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){
5558 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL);
5559 pT->wtFlags |= TERM_CODED;
5560 }
5561 }
5562
drh6457a352013-06-21 00:35:37 +00005563 if( wctrlFlags & WHERE_WANT_DISTINCT ){
drh9b3bfa02021-12-02 01:30:16 +00005564 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){
5565 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via
5566 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */
5567 wctrlFlags &= ~WHERE_WANT_DISTINCT;
5568 pWInfo->wctrlFlags &= ~WHERE_WANT_DISTINCT;
5569 }else if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){
drh6457a352013-06-21 00:35:37 +00005570 /* The DISTINCT marking is pointless. Ignore it. */
drh4f402f22013-06-11 18:59:38 +00005571 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
5572 }else if( pOrderBy==0 ){
drh6457a352013-06-21 00:35:37 +00005573 /* Try to ORDER BY the result set to make distinct processing easier */
drh4f402f22013-06-11 18:59:38 +00005574 pWInfo->wctrlFlags |= WHERE_DISTINCTBY;
drh6457a352013-06-21 00:35:37 +00005575 pWInfo->pOrderBy = pResultSet;
drh4f402f22013-06-11 18:59:38 +00005576 }
dan38cc40c2011-06-30 20:17:15 +00005577 }
5578
drhf1b5f5b2013-05-02 00:15:01 +00005579 /* Construct the WhereLoop objects */
drhc90713d2014-09-30 13:46:49 +00005580#if defined(WHERETRACE_ENABLED)
drhc3489bb2016-02-25 16:04:59 +00005581 if( sqlite3WhereTrace & 0xffff ){
5582 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags);
5583 if( wctrlFlags & WHERE_USE_LIMIT ){
5584 sqlite3DebugPrintf(", limit: %d", iAuxArg);
5585 }
5586 sqlite3DebugPrintf(")\n");
drh55b4c822019-08-03 16:17:46 +00005587 if( sqlite3WhereTrace & 0x100 ){
5588 Select sSelect;
5589 memset(&sSelect, 0, sizeof(sSelect));
5590 sSelect.selFlags = SF_WhereBegin;
5591 sSelect.pSrc = pTabList;
5592 sSelect.pWhere = pWhere;
5593 sSelect.pOrderBy = pOrderBy;
5594 sSelect.pEList = pResultSet;
5595 sqlite3TreeViewSelect(0, &sSelect, 0);
5596 }
drhc3489bb2016-02-25 16:04:59 +00005597 }
drhb121dd12015-06-06 18:30:17 +00005598 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */
drh05fbfd82019-12-05 17:31:58 +00005599 sqlite3DebugPrintf("---- WHERE clause at start of analysis:\n");
drhc84a4022016-05-27 12:30:20 +00005600 sqlite3WhereClausePrint(sWLB.pWC);
drhc90713d2014-09-30 13:46:49 +00005601 }
5602#endif
5603
drhb8a8e8a2013-06-10 19:12:39 +00005604 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){
drh60c96cd2013-06-09 17:21:25 +00005605 rc = whereLoopAddAll(&sWLB);
5606 if( rc ) goto whereBeginError;
drh89efac92020-02-22 16:58:49 +00005607
5608#ifdef SQLITE_ENABLE_STAT4
5609 /* If one or more WhereTerm.truthProb values were used in estimating
5610 ** loop parameters, but then those truthProb values were subsequently
5611 ** changed based on STAT4 information while computing subsequent loops,
5612 ** then we need to rerun the whole loop building process so that all
5613 ** loops will be built using the revised truthProb values. */
5614 if( sWLB.bldFlags2 & SQLITE_BLDF2_2NDPASS ){
drhcea19512020-02-22 18:27:48 +00005615 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC);
drh89efac92020-02-22 16:58:49 +00005616 WHERETRACE(0xffff,
drhf06cdde2020-02-24 16:46:08 +00005617 ("**** Redo all loop computations due to"
5618 " TERM_HIGHTRUTH changes ****\n"));
drh89efac92020-02-22 16:58:49 +00005619 while( pWInfo->pLoops ){
5620 WhereLoop *p = pWInfo->pLoops;
5621 pWInfo->pLoops = p->pNextLoop;
5622 whereLoopDelete(db, p);
5623 }
5624 rc = whereLoopAddAll(&sWLB);
5625 if( rc ) goto whereBeginError;
5626 }
5627#endif
drhcea19512020-02-22 18:27:48 +00005628 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC);
drh60c96cd2013-06-09 17:21:25 +00005629
drh4f402f22013-06-11 18:59:38 +00005630 wherePathSolver(pWInfo, 0);
drh60c96cd2013-06-09 17:21:25 +00005631 if( db->mallocFailed ) goto whereBeginError;
5632 if( pWInfo->pOrderBy ){
drhc7f0d222013-06-19 03:27:12 +00005633 wherePathSolver(pWInfo, pWInfo->nRowOut+1);
drh60c96cd2013-06-09 17:21:25 +00005634 if( db->mallocFailed ) goto whereBeginError;
drha18f3d22013-05-08 03:05:41 +00005635 }
5636 }
drh60c96cd2013-06-09 17:21:25 +00005637 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){
drh8426e362016-03-08 01:32:30 +00005638 pWInfo->revMask = ALLBITS;
drha50ef112013-05-22 02:06:59 +00005639 }
drh0c7d3d32022-01-24 16:47:12 +00005640 if( pParse->nErr ){
drh75b93402013-05-31 20:43:57 +00005641 goto whereBeginError;
5642 }
drh0c7d3d32022-01-24 16:47:12 +00005643 assert( db->mallocFailed==0 );
drhb121dd12015-06-06 18:30:17 +00005644#ifdef WHERETRACE_ENABLED
drha18f3d22013-05-08 03:05:41 +00005645 if( sqlite3WhereTrace ){
drh4f402f22013-06-11 18:59:38 +00005646 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut);
drhddba0c22014-03-18 20:33:42 +00005647 if( pWInfo->nOBSat>0 ){
5648 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask);
drh319f6772013-05-14 15:31:07 +00005649 }
drh4f402f22013-06-11 18:59:38 +00005650 switch( pWInfo->eDistinct ){
5651 case WHERE_DISTINCT_UNIQUE: {
5652 sqlite3DebugPrintf(" DISTINCT=unique");
5653 break;
5654 }
5655 case WHERE_DISTINCT_ORDERED: {
5656 sqlite3DebugPrintf(" DISTINCT=ordered");
5657 break;
5658 }
5659 case WHERE_DISTINCT_UNORDERED: {
5660 sqlite3DebugPrintf(" DISTINCT=unordered");
5661 break;
5662 }
5663 }
5664 sqlite3DebugPrintf("\n");
drhfd636c72013-06-21 02:05:06 +00005665 for(ii=0; ii<pWInfo->nLevel; ii++){
drhcacdf202019-12-28 13:39:47 +00005666 sqlite3WhereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC);
drhf1b5f5b2013-05-02 00:15:01 +00005667 }
5668 }
5669#endif
dan41203c62017-11-21 19:22:45 +00005670
drh70b403b2021-12-03 18:53:53 +00005671 /* Attempt to omit tables from a join that do not affect the result.
5672 ** See the comment on whereOmitNoopJoin() for further information.
dan41203c62017-11-21 19:22:45 +00005673 **
drh70b403b2021-12-03 18:53:53 +00005674 ** This query optimization is factored out into a separate "no-inline"
5675 ** procedure to keep the sqlite3WhereBegin() procedure from becoming
5676 ** too large. If sqlite3WhereBegin() becomes too large, that prevents
5677 ** some C-compiler optimizers from in-lining the
5678 ** sqlite3WhereCodeOneLoopStart() procedure, and it is important to
5679 ** in-line sqlite3WhereCodeOneLoopStart() for performance reasons.
dan41203c62017-11-21 19:22:45 +00005680 */
drh53bf7172017-11-23 04:45:35 +00005681 notReady = ~(Bitmask)0;
drh1031bd92013-06-22 15:44:26 +00005682 if( pWInfo->nLevel>=2
danf330d532021-04-03 19:23:59 +00005683 && pResultSet!=0 /* these two combine to guarantee */
5684 && 0==(wctrlFlags & WHERE_AGG_DISTINCT) /* condition (1) above */
drh1031bd92013-06-22 15:44:26 +00005685 && OptimizationEnabled(db, SQLITE_OmitNoopJoin)
5686 ){
drh70b403b2021-12-03 18:53:53 +00005687 notReady = whereOmitNoopJoin(pWInfo, notReady);
5688 nTabList = pWInfo->nLevel;
5689 assert( nTabList>0 );
drhfd636c72013-06-21 02:05:06 +00005690 }
drh70b403b2021-12-03 18:53:53 +00005691
drhfa35f5c2021-12-04 13:43:57 +00005692 /* Check to see if there are any SEARCH loops that might benefit from
5693 ** using a Bloom filter.
5694 */
5695 if( pWInfo->nLevel>=2
5696 && OptimizationEnabled(db, SQLITE_BloomFilter)
5697 ){
drhfecbf0a2021-12-04 21:11:18 +00005698 whereCheckIfBloomFilterIsUseful(pWInfo);
drhfa35f5c2021-12-04 13:43:57 +00005699 }
5700
drh05fbfd82019-12-05 17:31:58 +00005701#if defined(WHERETRACE_ENABLED)
5702 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */
5703 sqlite3DebugPrintf("---- WHERE clause at end of analysis:\n");
5704 sqlite3WhereClausePrint(sWLB.pWC);
5705 }
drh3b48e8c2013-06-12 20:18:16 +00005706 WHERETRACE(0xffff,("*** Optimizer Finished ***\n"));
drh05fbfd82019-12-05 17:31:58 +00005707#endif
drh8e23daf2013-06-11 13:30:04 +00005708 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut;
drhf1b5f5b2013-05-02 00:15:01 +00005709
drh08c88eb2008-04-10 13:33:18 +00005710 /* If the caller is an UPDATE or DELETE statement that is requesting
5711 ** to use a one-pass algorithm, determine if this is appropriate.
dan0c2ba132018-01-16 13:37:43 +00005712 **
5713 ** A one-pass approach can be used if the caller has requested one
5714 ** and either (a) the scan visits at most one row or (b) each
5715 ** of the following are true:
5716 **
5717 ** * the caller has indicated that a one-pass approach can be used
5718 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and
5719 ** * the table is not a virtual table, and
5720 ** * either the scan does not use the OR optimization or the caller
5721 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified
5722 ** for DELETE).
5723 **
5724 ** The last qualification is because an UPDATE statement uses
5725 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can
5726 ** use a one-pass approach, and this is not set accurately for scans
5727 ** that use the OR optimization.
drh08c88eb2008-04-10 13:33:18 +00005728 */
drh165be382008-12-05 02:36:33 +00005729 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 );
danf0ee1d32015-09-12 19:26:11 +00005730 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){
5731 int wsFlags = pWInfo->a[0].pWLoop->wsFlags;
5732 int bOnerow = (wsFlags & WHERE_ONEROW)!=0;
dan58ed3742019-01-15 14:31:01 +00005733 assert( !(wsFlags & WHERE_VIRTUALTABLE) || IsVirtual(pTabList->a[0].pTab) );
dan0c2ba132018-01-16 13:37:43 +00005734 if( bOnerow || (
5735 0!=(wctrlFlags & WHERE_ONEPASS_MULTIROW)
dan58ed3742019-01-15 14:31:01 +00005736 && !IsVirtual(pTabList->a[0].pTab)
dan0c2ba132018-01-16 13:37:43 +00005737 && (0==(wsFlags & WHERE_MULTI_OR) || (wctrlFlags & WHERE_DUPLICATES_OK))
5738 )){
drhb0264ee2015-09-14 14:45:50 +00005739 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI;
danfd261ec2015-10-22 20:54:33 +00005740 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){
5741 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){
5742 bFordelete = OPFLAG_FORDELETE;
5743 }
5744 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY);
danf0ee1d32015-09-12 19:26:11 +00005745 }
drh702ba9f2013-11-07 21:25:13 +00005746 }
drh08c88eb2008-04-10 13:33:18 +00005747 }
drheb04de32013-05-10 15:16:30 +00005748
drh9012bcb2004-12-19 00:11:35 +00005749 /* Open all tables in the pTabList and any indices selected for
5750 ** searching those tables.
5751 */
drh9cd1c992012-09-25 20:43:35 +00005752 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){
danielk1977da184232006-01-05 11:34:32 +00005753 Table *pTab; /* Table to open */
danielk1977da184232006-01-05 11:34:32 +00005754 int iDb; /* Index of database containing table/index */
drh76012942021-02-21 21:04:54 +00005755 SrcItem *pTabItem;
drh9012bcb2004-12-19 00:11:35 +00005756
drh29dda4a2005-07-21 18:23:20 +00005757 pTabItem = &pTabList->a[pLevel->iFrom];
drh9012bcb2004-12-19 00:11:35 +00005758 pTab = pTabItem->pTab;
danielk1977595a5232009-07-24 17:58:53 +00005759 iDb = sqlite3SchemaToIndex(db, pTab->pSchema);
drh7ba39a92013-05-30 17:43:19 +00005760 pLoop = pLevel->pWLoop;
drhf38524d2021-08-02 16:41:57 +00005761 if( (pTab->tabFlags & TF_Ephemeral)!=0 || IsView(pTab) ){
drh75bb9f52010-04-06 18:51:42 +00005762 /* Do nothing */
5763 }else
drh9eff6162006-06-12 21:59:13 +00005764#ifndef SQLITE_OMIT_VIRTUALTABLE
drh7ba39a92013-05-30 17:43:19 +00005765 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){
danielk1977595a5232009-07-24 17:58:53 +00005766 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab);
danielk197793626f42006-06-20 13:07:27 +00005767 int iCur = pTabItem->iCursor;
danielk1977595a5232009-07-24 17:58:53 +00005768 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB);
drhfc5e5462012-12-03 17:04:40 +00005769 }else if( IsVirtual(pTab) ){
5770 /* noop */
drh9eff6162006-06-12 21:59:13 +00005771 }else
5772#endif
drha20c71e2022-04-10 16:13:37 +00005773 if( ((pLoop->wsFlags & WHERE_IDX_ONLY)==0
5774 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0)
5775 || (pTabItem->fg.jointype & (JT_LTORJ|JT_RIGHT))!=0
5776 ){
drhfc8d4f92013-11-08 15:19:46 +00005777 int op = OP_OpenRead;
drhb0264ee2015-09-14 14:45:50 +00005778 if( pWInfo->eOnePass!=ONEPASS_OFF ){
drhfc8d4f92013-11-08 15:19:46 +00005779 op = OP_OpenWrite;
5780 pWInfo->aiCurOnePass[0] = pTabItem->iCursor;
5781 };
drh08c88eb2008-04-10 13:33:18 +00005782 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op);
drhfc8d4f92013-11-08 15:19:46 +00005783 assert( pTabItem->iCursor==pLevel->iTabCur );
drhb0264ee2015-09-14 14:45:50 +00005784 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 );
5785 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS );
drh1a9082f2019-11-01 15:19:24 +00005786 if( pWInfo->eOnePass==ONEPASS_OFF
5787 && pTab->nCol<BMS
5788 && (pTab->tabFlags & (TF_HasGenerated|TF_WithoutRowid))==0
drh87fb37e2022-01-31 15:59:43 +00005789 && (pLoop->wsFlags & (WHERE_AUTO_INDEX|WHERE_BLOOMFILTER))==0
drh1a9082f2019-11-01 15:19:24 +00005790 ){
5791 /* If we know that only a prefix of the record will be used,
5792 ** it is advantageous to reduce the "column count" field in
5793 ** the P4 operand of the OP_OpenRead/Write opcode. */
danielk19779792eef2006-01-13 15:58:43 +00005794 Bitmask b = pTabItem->colUsed;
5795 int n = 0;
drh74161702006-02-24 02:53:49 +00005796 for(; b; b=b>>1, n++){}
drh00dceca2016-01-11 22:58:50 +00005797 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32);
danielk19779792eef2006-01-13 15:58:43 +00005798 assert( n<=pTab->nCol );
5799 }
drhba25c7e2020-03-12 17:54:39 +00005800#ifdef SQLITE_ENABLE_CURSOR_HINTS
danc5dc3dc2015-10-26 20:11:24 +00005801 if( pLoop->u.btree.pIndex!=0 ){
5802 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete);
5803 }else
drh2f2b0272015-08-14 18:50:04 +00005804#endif
danc5dc3dc2015-10-26 20:11:24 +00005805 {
5806 sqlite3VdbeChangeP5(v, bFordelete);
5807 }
drh97bae792015-06-05 15:59:57 +00005808#ifdef SQLITE_ENABLE_COLUMN_USED_MASK
5809 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0,
5810 (const u8*)&pTabItem->colUsed, P4_INT64);
5811#endif
danielk1977c00da102006-01-07 13:21:04 +00005812 }else{
5813 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName);
drh9012bcb2004-12-19 00:11:35 +00005814 }
drh7e47cb82013-05-31 17:55:27 +00005815 if( pLoop->wsFlags & WHERE_INDEXED ){
drh7ba39a92013-05-30 17:43:19 +00005816 Index *pIx = pLoop->u.btree.pIndex;
drhfc8d4f92013-11-08 15:19:46 +00005817 int iIndexCur;
5818 int op = OP_OpenRead;
drh154896e2017-09-15 14:36:13 +00005819 /* iAuxArg is always set to a positive value if ONEPASS is possible */
drhc3489bb2016-02-25 16:04:59 +00005820 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 );
drh48dd1d82014-05-27 18:18:58 +00005821 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx)
drhce943bc2016-05-19 18:56:33 +00005822 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0
drha3bc66a2014-05-27 17:57:32 +00005823 ){
5824 /* This is one term of an OR-optimization using the PRIMARY KEY of a
5825 ** WITHOUT ROWID table. No need for a separate index */
5826 iIndexCur = pLevel->iTabCur;
5827 op = 0;
drhb0264ee2015-09-14 14:45:50 +00005828 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){
drhfc8d4f92013-11-08 15:19:46 +00005829 Index *pJ = pTabItem->pTab->pIndex;
drhc3489bb2016-02-25 16:04:59 +00005830 iIndexCur = iAuxArg;
drhfc8d4f92013-11-08 15:19:46 +00005831 assert( wctrlFlags & WHERE_ONEPASS_DESIRED );
5832 while( ALWAYS(pJ) && pJ!=pIx ){
5833 iIndexCur++;
5834 pJ = pJ->pNext;
5835 }
5836 op = OP_OpenWrite;
5837 pWInfo->aiCurOnePass[1] = iIndexCur;
drhce943bc2016-05-19 18:56:33 +00005838 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){
drhc3489bb2016-02-25 16:04:59 +00005839 iIndexCur = iAuxArg;
drha72a15e2016-05-09 19:58:56 +00005840 op = OP_ReopenIdx;
drhfc8d4f92013-11-08 15:19:46 +00005841 }else{
5842 iIndexCur = pParse->nTab++;
5843 }
5844 pLevel->iIdxCur = iIndexCur;
danielk1977da184232006-01-05 11:34:32 +00005845 assert( pIx->pSchema==pTab->pSchema );
drhb0367fb2012-08-25 02:11:13 +00005846 assert( iIndexCur>=0 );
drha3bc66a2014-05-27 17:57:32 +00005847 if( op ){
5848 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb);
5849 sqlite3VdbeSetP4KeyInfo(pParse, pIx);
drhe0997b32015-03-20 14:57:50 +00005850 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0
5851 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0
dan15750a22019-08-16 21:07:19 +00005852 && (pLoop->wsFlags & WHERE_BIGNULL_SORT)==0
drh68cf0ac2020-09-28 19:51:54 +00005853 && (pLoop->wsFlags & WHERE_IN_SEEKSCAN)==0
drhe0997b32015-03-20 14:57:50 +00005854 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0
drh8489bf52017-04-13 01:19:30 +00005855 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED
drhe0997b32015-03-20 14:57:50 +00005856 ){
drh576d0a92020-03-12 17:28:27 +00005857 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ);
drhe0997b32015-03-20 14:57:50 +00005858 }
drha3bc66a2014-05-27 17:57:32 +00005859 VdbeComment((v, "%s", pIx->zName));
drh97bae792015-06-05 15:59:57 +00005860#ifdef SQLITE_ENABLE_COLUMN_USED_MASK
5861 {
5862 u64 colUsed = 0;
5863 int ii, jj;
5864 for(ii=0; ii<pIx->nColumn; ii++){
5865 jj = pIx->aiColumn[ii];
5866 if( jj<0 ) continue;
5867 if( jj>63 ) jj = 63;
5868 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue;
5869 colUsed |= ((u64)1)<<(ii<63 ? ii : 63);
5870 }
5871 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0,
5872 (u8*)&colUsed, P4_INT64);
5873 }
5874#endif /* SQLITE_ENABLE_COLUMN_USED_MASK */
drha3bc66a2014-05-27 17:57:32 +00005875 }
drh9012bcb2004-12-19 00:11:35 +00005876 }
drhaceb31b2014-02-08 01:40:27 +00005877 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb);
drh7c1734b2022-04-09 12:27:20 +00005878 if( (pTabItem->fg.jointype & JT_RIGHT)!=0
5879 && (pLevel->pRJ = sqlite3WhereMalloc(pWInfo, sizeof(WhereRightJoin)))!=0
5880 ){
5881 WhereRightJoin *pRJ = pLevel->pRJ;
5882 pRJ->iMatch = pParse->nTab++;
5883 pRJ->regBloom = ++pParse->nMem;
5884 sqlite3VdbeAddOp2(v, OP_Blob, 65536, pRJ->regBloom);
5885 pRJ->regReturn = ++pParse->nMem;
drh2bd9f442022-04-17 20:30:52 +00005886 sqlite3VdbeAddOp2(v, OP_Null, 0, pRJ->regReturn);
drhc2308ad2022-04-09 03:16:26 +00005887 assert( pTab==pTabItem->pTab );
drhc2308ad2022-04-09 03:16:26 +00005888 if( HasRowid(pTab) ){
5889 KeyInfo *pInfo;
drh7c1734b2022-04-09 12:27:20 +00005890 sqlite3VdbeAddOp2(v, OP_OpenEphemeral, pRJ->iMatch, 1);
drhc2308ad2022-04-09 03:16:26 +00005891 pInfo = sqlite3KeyInfoAlloc(pParse->db, 1, 0);
5892 if( pInfo ){
5893 pInfo->aColl[0] = 0;
5894 pInfo->aSortFlags[0] = 0;
5895 sqlite3VdbeAppendP4(v, pInfo, P4_KEYINFO);
5896 }
5897 }else{
5898 Index *pPk = sqlite3PrimaryKeyIndex(pTab);
drh7c1734b2022-04-09 12:27:20 +00005899 sqlite3VdbeAddOp2(v, OP_OpenEphemeral, pRJ->iMatch, pPk->nKeyCol);
drhc2308ad2022-04-09 03:16:26 +00005900 sqlite3VdbeSetP4KeyInfo(pParse, pPk);
5901 }
drhe21e36d2022-04-10 17:14:48 +00005902 pLoop->wsFlags &= ~WHERE_IDX_ONLY;
drh12c35ec2022-04-13 12:12:01 +00005903 /* The nature of RIGHT JOIN processing is such that it messes up
5904 ** the output order. So omit any ORDER BY/GROUP BY elimination
5905 ** optimizations. We need to do an actual sort for RIGHT JOIN. */
5906 pWInfo->nOBSat = 0;
5907 pWInfo->eDistinct = WHERE_DISTINCT_UNORDERED;
drha76ac882022-04-08 19:20:12 +00005908 }
drh9012bcb2004-12-19 00:11:35 +00005909 }
5910 pWInfo->iTop = sqlite3VdbeCurrentAddr(v);
drha21a64d2010-04-06 22:33:55 +00005911 if( db->mallocFailed ) goto whereBeginError;
drh9012bcb2004-12-19 00:11:35 +00005912
drh29dda4a2005-07-21 18:23:20 +00005913 /* Generate the code to do the search. Each iteration of the for
5914 ** loop below generates code for a single nested loop of the VM
5915 ** program.
drh75897232000-05-29 14:26:00 +00005916 */
drh9cd1c992012-09-25 20:43:35 +00005917 for(ii=0; ii<nTabList; ii++){
dan6f9702e2014-11-01 20:38:06 +00005918 int addrExplain;
5919 int wsFlags;
drh6d64b4a2021-11-07 23:33:01 +00005920 if( pParse->nErr ) goto whereBeginError;
drh9cd1c992012-09-25 20:43:35 +00005921 pLevel = &pWInfo->a[ii];
dan6f9702e2014-11-01 20:38:06 +00005922 wsFlags = pLevel->pWLoop->wsFlags;
drhfa35f5c2021-12-04 13:43:57 +00005923 if( (wsFlags & (WHERE_AUTO_INDEX|WHERE_BLOOMFILTER))!=0 ){
5924 if( (wsFlags & WHERE_AUTO_INDEX)!=0 ){
drhcc04afd2013-08-22 02:56:28 +00005925#ifndef SQLITE_OMIT_AUTOMATIC_INDEX
drhfa35f5c2021-12-04 13:43:57 +00005926 constructAutomaticIndex(pParse, &pWInfo->sWC,
5927 &pTabList->a[pLevel->iFrom], notReady, pLevel);
5928#endif
5929 }else{
drh27a9e1f2021-12-10 17:36:16 +00005930 sqlite3ConstructBloomFilter(pWInfo, ii, pLevel, notReady);
drhfa35f5c2021-12-04 13:43:57 +00005931 }
drhcc04afd2013-08-22 02:56:28 +00005932 if( db->mallocFailed ) goto whereBeginError;
5933 }
drh6f82e852015-06-06 20:12:09 +00005934 addrExplain = sqlite3WhereExplainOneScan(
drhe2188f02018-05-07 11:37:34 +00005935 pParse, pTabList, pLevel, wctrlFlags
dan6f9702e2014-11-01 20:38:06 +00005936 );
drhcc04afd2013-08-22 02:56:28 +00005937 pLevel->addrBody = sqlite3VdbeCurrentAddr(v);
drh47df8a22018-12-25 00:15:37 +00005938 notReady = sqlite3WhereCodeOneLoopStart(pParse,v,pWInfo,ii,pLevel,notReady);
dan4a07e3d2010-11-09 14:48:59 +00005939 pWInfo->iContinue = pLevel->addrCont;
drhce943bc2016-05-19 18:56:33 +00005940 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){
drh6f82e852015-06-06 20:12:09 +00005941 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain);
dan6f9702e2014-11-01 20:38:06 +00005942 }
drh75897232000-05-29 14:26:00 +00005943 }
drh7ec764a2005-07-21 03:48:20 +00005944
drh6fa978d2013-05-30 19:29:19 +00005945 /* Done. */
drh6bc69a22013-11-19 12:33:23 +00005946 VdbeModuleComment((v, "Begin WHERE-core"));
drh5e6d90f2020-08-14 17:39:31 +00005947 pWInfo->iEndWhere = sqlite3VdbeCurrentAddr(v);
drh75897232000-05-29 14:26:00 +00005948 return pWInfo;
drhe23399f2005-07-22 00:31:39 +00005949
5950 /* Jump here if malloc fails */
danielk197785574e32008-10-06 05:32:18 +00005951whereBeginError:
drh8b307fb2010-04-06 15:57:05 +00005952 if( pWInfo ){
drhd784cc82021-04-15 12:56:44 +00005953 testcase( pWInfo->pExprMods!=0 );
5954 whereUndoExprMods(pWInfo);
drh8b307fb2010-04-06 15:57:05 +00005955 pParse->nQueryLoop = pWInfo->savedNQueryLoop;
5956 whereInfoFree(db, pWInfo);
5957 }
drhe23399f2005-07-22 00:31:39 +00005958 return 0;
drh75897232000-05-29 14:26:00 +00005959}
5960
5961/*
drh299bf7c2018-06-11 17:35:02 +00005962** Part of sqlite3WhereEnd() will rewrite opcodes to reference the
5963** index rather than the main table. In SQLITE_DEBUG mode, we want
5964** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine
5965** does that.
5966*/
5967#ifndef SQLITE_DEBUG
5968# define OpcodeRewriteTrace(D,K,P) /* no-op */
5969#else
5970# define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P)
5971 static void sqlite3WhereOpcodeRewriteTrace(
5972 sqlite3 *db,
5973 int pc,
5974 VdbeOp *pOp
5975 ){
5976 if( (db->flags & SQLITE_VdbeAddopTrace)==0 ) return;
5977 sqlite3VdbePrintOp(0, pc, pOp);
5978 }
5979#endif
5980
drh75c493f2022-03-21 13:47:15 +00005981#ifdef SQLITE_DEBUG
5982/*
5983** Return true if cursor iCur is opened by instruction k of the
5984** bytecode. Used inside of assert() only.
5985*/
5986static int cursorIsOpen(Vdbe *v, int iCur, int k){
5987 while( k>=0 ){
5988 VdbeOp *pOp = sqlite3VdbeGetOp(v,k--);
5989 if( pOp->p1!=iCur ) continue;
5990 if( pOp->opcode==OP_Close ) return 0;
5991 if( pOp->opcode==OP_OpenRead ) return 1;
5992 if( pOp->opcode==OP_OpenWrite ) return 1;
5993 if( pOp->opcode==OP_OpenDup ) return 1;
5994 if( pOp->opcode==OP_OpenAutoindex ) return 1;
5995 if( pOp->opcode==OP_OpenEphemeral ) return 1;
5996 }
5997 return 0;
5998}
5999#endif /* SQLITE_DEBUG */
6000
drh299bf7c2018-06-11 17:35:02 +00006001/*
drhc27a1ce2002-06-14 20:58:45 +00006002** Generate the end of the WHERE loop. See comments on
danielk19774adee202004-05-08 08:23:19 +00006003** sqlite3WhereBegin() for additional information.
drh75897232000-05-29 14:26:00 +00006004*/
danielk19774adee202004-05-08 08:23:19 +00006005void sqlite3WhereEnd(WhereInfo *pWInfo){
drh633e6d52008-07-28 19:34:53 +00006006 Parse *pParse = pWInfo->pParse;
6007 Vdbe *v = pParse->pVdbe;
drh19a775c2000-06-05 18:54:46 +00006008 int i;
drh6b563442001-11-07 16:48:26 +00006009 WhereLevel *pLevel;
drh7ba39a92013-05-30 17:43:19 +00006010 WhereLoop *pLoop;
drhad3cab52002-05-24 02:04:32 +00006011 SrcList *pTabList = pWInfo->pTabList;
drh633e6d52008-07-28 19:34:53 +00006012 sqlite3 *db = pParse->db;
drhf8556d02020-08-14 21:32:16 +00006013 int iEnd = sqlite3VdbeCurrentAddr(v);
drh19a775c2000-06-05 18:54:46 +00006014
drh9012bcb2004-12-19 00:11:35 +00006015 /* Generate loop termination code.
6016 */
drh6bc69a22013-11-19 12:33:23 +00006017 VdbeModuleComment((v, "End WHERE-core"));
drhc01a3c12009-12-16 22:10:49 +00006018 for(i=pWInfo->nLevel-1; i>=0; i--){
drhcd8629e2013-11-13 12:27:25 +00006019 int addr;
drh6b563442001-11-07 16:48:26 +00006020 pLevel = &pWInfo->a[i];
drh7c1734b2022-04-09 12:27:20 +00006021 if( pLevel->pRJ ){
drh86c1beb2022-04-09 14:48:35 +00006022 /* Terminate the subroutine that forms the interior of the loop of
6023 ** the RIGHT JOIN table */
6024 WhereRightJoin *pRJ = pLevel->pRJ;
drh6134b2d2022-04-11 17:27:38 +00006025 sqlite3VdbeResolveLabel(v, pLevel->addrCont);
6026 pLevel->addrCont = 0;
drhb77c3122022-04-23 18:04:31 +00006027 pRJ->endSubrtn = sqlite3VdbeCurrentAddr(v);
drh2bd9f442022-04-17 20:30:52 +00006028 sqlite3VdbeAddOp3(v, OP_Return, pRJ->regReturn, pRJ->addrSubrtn, 1);
6029 VdbeCoverage(v);
drh2c31c002022-04-14 16:34:07 +00006030 assert( pParse->withinRJSubrtn>0 );
6031 pParse->withinRJSubrtn--;
drh7c1734b2022-04-09 12:27:20 +00006032 }
drh7ba39a92013-05-30 17:43:19 +00006033 pLoop = pLevel->pWLoop;
drh6b563442001-11-07 16:48:26 +00006034 if( pLevel->op!=OP_Noop ){
drh8489bf52017-04-13 01:19:30 +00006035#ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT
drhc04ea802017-04-13 19:48:29 +00006036 int addrSeek = 0;
drh839fa6d2017-04-13 13:01:59 +00006037 Index *pIdx;
drh172806e2017-04-13 21:29:02 +00006038 int n;
drh8489bf52017-04-13 01:19:30 +00006039 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED
drhfa337cc2017-11-23 00:45:21 +00006040 && i==pWInfo->nLevel-1 /* Ticket [ef9318757b152e3] 2017-10-21 */
drh8489bf52017-04-13 01:19:30 +00006041 && (pLoop->wsFlags & WHERE_INDEXED)!=0
drh839fa6d2017-04-13 13:01:59 +00006042 && (pIdx = pLoop->u.btree.pIndex)->hasStat1
dana79a0e72019-07-29 14:42:56 +00006043 && (n = pLoop->u.btree.nDistinctCol)>0
drha2e2d922017-04-14 22:41:27 +00006044 && pIdx->aiRowLogEst[n]>=36
drh8489bf52017-04-13 01:19:30 +00006045 ){
drh172806e2017-04-13 21:29:02 +00006046 int r1 = pParse->nMem+1;
6047 int j, op;
drh8489bf52017-04-13 01:19:30 +00006048 for(j=0; j<n; j++){
6049 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j);
6050 }
drh172806e2017-04-13 21:29:02 +00006051 pParse->nMem += n+1;
drh8489bf52017-04-13 01:19:30 +00006052 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT;
drhc04ea802017-04-13 19:48:29 +00006053 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n);
drh8489bf52017-04-13 01:19:30 +00006054 VdbeCoverageIf(v, op==OP_SeekLT);
6055 VdbeCoverageIf(v, op==OP_SeekGT);
6056 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2);
drh8489bf52017-04-13 01:19:30 +00006057 }
drhc04ea802017-04-13 19:48:29 +00006058#endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */
6059 /* The common case: Advance to the next row */
drh6134b2d2022-04-11 17:27:38 +00006060 if( pLevel->addrCont ) sqlite3VdbeResolveLabel(v, pLevel->addrCont);
drhe39a7322014-02-03 14:04:11 +00006061 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3);
drhd1d38482008-10-07 23:46:38 +00006062 sqlite3VdbeChangeP5(v, pLevel->p5);
drh688852a2014-02-17 22:40:43 +00006063 VdbeCoverage(v);
drh7d176102014-02-18 03:07:12 +00006064 VdbeCoverageIf(v, pLevel->op==OP_Next);
6065 VdbeCoverageIf(v, pLevel->op==OP_Prev);
6066 VdbeCoverageIf(v, pLevel->op==OP_VNext);
dan15750a22019-08-16 21:07:19 +00006067 if( pLevel->regBignull ){
6068 sqlite3VdbeResolveLabel(v, pLevel->addrBignull);
danbd717a42019-08-29 21:16:46 +00006069 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, pLevel->regBignull, pLevel->p2-1);
drhdb586e42019-08-29 16:48:10 +00006070 VdbeCoverage(v);
dan15750a22019-08-16 21:07:19 +00006071 }
drhc04ea802017-04-13 19:48:29 +00006072#ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT
6073 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek);
6074#endif
drh6134b2d2022-04-11 17:27:38 +00006075 }else if( pLevel->addrCont ){
dana74f5c22017-04-13 18:33:33 +00006076 sqlite3VdbeResolveLabel(v, pLevel->addrCont);
drh19a775c2000-06-05 18:54:46 +00006077 }
drh04756292021-10-14 19:28:28 +00006078 if( (pLoop->wsFlags & WHERE_IN_ABLE)!=0 && pLevel->u.in.nIn>0 ){
drh72e8fa42007-03-28 14:30:06 +00006079 struct InLoop *pIn;
drhe23399f2005-07-22 00:31:39 +00006080 int j;
drhb3190c12008-12-08 21:37:14 +00006081 sqlite3VdbeResolveLabel(v, pLevel->addrNxt);
drh111a6a72008-12-21 03:51:16 +00006082 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){
drh81f5ef02021-04-29 15:49:34 +00006083 assert( sqlite3VdbeGetOp(v, pIn->addrInTop+1)->opcode==OP_IsNull
6084 || pParse->db->mallocFailed );
drhb3190c12008-12-08 21:37:14 +00006085 sqlite3VdbeJumpHere(v, pIn->addrInTop+1);
dan8da209b2016-07-26 18:06:08 +00006086 if( pIn->eEndLoopOp!=OP_Noop ){
drha0368d92018-05-30 00:54:23 +00006087 if( pIn->nPrefix ){
drhf761d932020-09-29 01:48:46 +00006088 int bEarlyOut =
6089 (pLoop->wsFlags & WHERE_VIRTUALTABLE)==0
6090 && (pLoop->wsFlags & WHERE_IN_EARLYOUT)!=0;
dan74ebaad2020-01-04 16:55:57 +00006091 if( pLevel->iLeftJoin ){
6092 /* For LEFT JOIN queries, cursor pIn->iCur may not have been
6093 ** opened yet. This occurs for WHERE clauses such as
6094 ** "a = ? AND b IN (...)", where the index is on (a, b). If
6095 ** the RHS of the (a=?) is NULL, then the "b IN (...)" may
6096 ** never have been coded, but the body of the loop run to
6097 ** return the null-row. So, if the cursor is not open yet,
6098 ** jump over the OP_Next or OP_Prev instruction about to
6099 ** be coded. */
6100 sqlite3VdbeAddOp2(v, OP_IfNotOpen, pIn->iCur,
drhf761d932020-09-29 01:48:46 +00006101 sqlite3VdbeCurrentAddr(v) + 2 + bEarlyOut);
dan74ebaad2020-01-04 16:55:57 +00006102 VdbeCoverage(v);
6103 }
drhf761d932020-09-29 01:48:46 +00006104 if( bEarlyOut ){
drh14c98a42020-03-16 03:07:53 +00006105 sqlite3VdbeAddOp4Int(v, OP_IfNoHope, pLevel->iIdxCur,
6106 sqlite3VdbeCurrentAddr(v)+2,
6107 pIn->iBase, pIn->nPrefix);
6108 VdbeCoverage(v);
drh81f5ef02021-04-29 15:49:34 +00006109 /* Retarget the OP_IsNull against the left operand of IN so
6110 ** it jumps past the OP_IfNoHope. This is because the
6111 ** OP_IsNull also bypasses the OP_Affinity opcode that is
6112 ** required by OP_IfNoHope. */
6113 sqlite3VdbeJumpHere(v, pIn->addrInTop+1);
drh14c98a42020-03-16 03:07:53 +00006114 }
drha0368d92018-05-30 00:54:23 +00006115 }
dan8da209b2016-07-26 18:06:08 +00006116 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop);
6117 VdbeCoverage(v);
drhf1949b62018-06-07 17:32:59 +00006118 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Prev);
6119 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Next);
dan8da209b2016-07-26 18:06:08 +00006120 }
drhb3190c12008-12-08 21:37:14 +00006121 sqlite3VdbeJumpHere(v, pIn->addrInTop-1);
drhe23399f2005-07-22 00:31:39 +00006122 }
drhd99f7062002-06-08 23:25:08 +00006123 }
drhb3190c12008-12-08 21:37:14 +00006124 sqlite3VdbeResolveLabel(v, pLevel->addrBrk);
drhf68621f2022-04-17 22:58:23 +00006125 if( pLevel->pRJ ){
6126 sqlite3VdbeAddOp3(v, OP_Return, pLevel->pRJ->regReturn, 0, 1);
6127 VdbeCoverage(v);
6128 }
drhcd8629e2013-11-13 12:27:25 +00006129 if( pLevel->addrSkip ){
drh076e85f2015-09-03 13:46:12 +00006130 sqlite3VdbeGoto(v, pLevel->addrSkip);
drhe084f402013-11-13 17:24:38 +00006131 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName));
drh2e5ef4e2013-11-13 16:58:54 +00006132 sqlite3VdbeJumpHere(v, pLevel->addrSkip);
6133 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2);
drhcd8629e2013-11-13 12:27:25 +00006134 }
drh41d2e662015-12-01 21:23:07 +00006135#ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS
drhf07cf6e2015-03-06 16:45:16 +00006136 if( pLevel->addrLikeRep ){
drh44aebff2016-05-02 10:25:42 +00006137 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1),
6138 pLevel->addrLikeRep);
drhf07cf6e2015-03-06 16:45:16 +00006139 VdbeCoverage(v);
drhf07cf6e2015-03-06 16:45:16 +00006140 }
drh41d2e662015-12-01 21:23:07 +00006141#endif
drhad2d8302002-05-24 20:31:36 +00006142 if( pLevel->iLeftJoin ){
danb40897a2016-10-26 15:46:09 +00006143 int ws = pLoop->wsFlags;
drh688852a2014-02-17 22:40:43 +00006144 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v);
danb40897a2016-10-26 15:46:09 +00006145 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 );
6146 if( (ws & WHERE_IDX_ONLY)==0 ){
dan41203c62017-11-21 19:22:45 +00006147 assert( pLevel->iTabCur==pTabList->a[pLevel->iFrom].iCursor );
6148 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iTabCur);
drh35451c62009-11-12 04:26:39 +00006149 }
danb40897a2016-10-26 15:46:09 +00006150 if( (ws & WHERE_INDEXED)
drh04756292021-10-14 19:28:28 +00006151 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCoveringIdx)
danb40897a2016-10-26 15:46:09 +00006152 ){
drh415ac682021-06-22 23:24:58 +00006153 if( ws & WHERE_MULTI_OR ){
drh04756292021-10-14 19:28:28 +00006154 Index *pIx = pLevel->u.pCoveringIdx;
drh415ac682021-06-22 23:24:58 +00006155 int iDb = sqlite3SchemaToIndex(db, pIx->pSchema);
6156 sqlite3VdbeAddOp3(v, OP_ReopenIdx, pLevel->iIdxCur, pIx->tnum, iDb);
6157 sqlite3VdbeSetP4KeyInfo(pParse, pIx);
6158 }
drh3c84ddf2008-01-09 02:15:38 +00006159 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur);
drh7f09b3e2002-08-13 13:15:49 +00006160 }
drh336a5302009-04-24 15:46:21 +00006161 if( pLevel->op==OP_Return ){
6162 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst);
6163 }else{
drh076e85f2015-09-03 13:46:12 +00006164 sqlite3VdbeGoto(v, pLevel->addrFirst);
drh336a5302009-04-24 15:46:21 +00006165 }
drhd654be82005-09-20 17:42:23 +00006166 sqlite3VdbeJumpHere(v, addr);
drhad2d8302002-05-24 20:31:36 +00006167 }
drh6bc69a22013-11-19 12:33:23 +00006168 VdbeModuleComment((v, "End WHERE-loop%d: %s", i,
drhfc8d4f92013-11-08 15:19:46 +00006169 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName));
drh19a775c2000-06-05 18:54:46 +00006170 }
drh9012bcb2004-12-19 00:11:35 +00006171
drhfd636c72013-06-21 02:05:06 +00006172 assert( pWInfo->nLevel<=pTabList->nSrc );
drhc01a3c12009-12-16 22:10:49 +00006173 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){
drh5f612292014-02-08 23:20:32 +00006174 int k, last;
drhf8556d02020-08-14 21:32:16 +00006175 VdbeOp *pOp, *pLastOp;
danbfca6a42012-08-24 10:52:35 +00006176 Index *pIdx = 0;
drh76012942021-02-21 21:04:54 +00006177 SrcItem *pTabItem = &pTabList->a[pLevel->iFrom];
drh9012bcb2004-12-19 00:11:35 +00006178 Table *pTab = pTabItem->pTab;
drh5cf590c2003-04-24 01:45:04 +00006179 assert( pTab!=0 );
drh7ba39a92013-05-30 17:43:19 +00006180 pLoop = pLevel->pWLoop;
drhfc8d4f92013-11-08 15:19:46 +00006181
drha76ac882022-04-08 19:20:12 +00006182 /* Do RIGHT JOIN processing. Generate code that will output the
6183 ** unmatched rows of the right operand of the RIGHT JOIN with
6184 ** all of the columns of the left operand set to NULL.
6185 */
drh86c1beb2022-04-09 14:48:35 +00006186 if( pLevel->pRJ ){
drh949e2ab2022-04-12 18:04:29 +00006187 sqlite3WhereRightJoinLoop(pWInfo, i, pLevel);
drhe21e36d2022-04-10 17:14:48 +00006188 continue;
drha76ac882022-04-08 19:20:12 +00006189 }
6190
drh5f612292014-02-08 23:20:32 +00006191 /* For a co-routine, change all OP_Column references to the table of
drh7b3aa082015-05-29 13:55:33 +00006192 ** the co-routine into OP_Copy of result contained in a register.
drh5f612292014-02-08 23:20:32 +00006193 ** OP_Rowid becomes OP_Null.
6194 */
drh202230e2017-03-11 13:02:59 +00006195 if( pTabItem->fg.viaCoroutine ){
6196 testcase( pParse->db->mallocFailed );
6197 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur,
danfb785b22015-10-24 20:31:22 +00006198 pTabItem->regResult, 0);
drh5f612292014-02-08 23:20:32 +00006199 continue;
6200 }
6201
drhf0030762013-06-14 13:27:01 +00006202 /* If this scan uses an index, make VDBE code substitutions to read data
6203 ** from the index instead of from the table where possible. In some cases
6204 ** this optimization prevents the table from ever being read, which can
6205 ** yield a significant performance boost.
drh9012bcb2004-12-19 00:11:35 +00006206 **
6207 ** Calls to the code generator in between sqlite3WhereBegin and
6208 ** sqlite3WhereEnd will have created code that references the table
6209 ** directly. This loop scans all that code looking for opcodes
6210 ** that reference the table and converts them into opcodes that
6211 ** reference the index.
6212 */
drh7ba39a92013-05-30 17:43:19 +00006213 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){
6214 pIdx = pLoop->u.btree.pIndex;
6215 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){
drh04756292021-10-14 19:28:28 +00006216 pIdx = pLevel->u.pCoveringIdx;
danbfca6a42012-08-24 10:52:35 +00006217 }
drh63c85a72015-09-28 14:40:20 +00006218 if( pIdx
drh63c85a72015-09-28 14:40:20 +00006219 && !db->mallocFailed
6220 ){
drh5e6d90f2020-08-14 17:39:31 +00006221 if( pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable) ){
drhf8556d02020-08-14 21:32:16 +00006222 last = iEnd;
drh5e6d90f2020-08-14 17:39:31 +00006223 }else{
6224 last = pWInfo->iEndWhere;
6225 }
drhf8556d02020-08-14 21:32:16 +00006226 k = pLevel->addrBody + 1;
drh299bf7c2018-06-11 17:35:02 +00006227#ifdef SQLITE_DEBUG
6228 if( db->flags & SQLITE_VdbeAddopTrace ){
6229 printf("TRANSLATE opcodes in range %d..%d\n", k, last-1);
6230 }
drhf8556d02020-08-14 21:32:16 +00006231 /* Proof that the "+1" on the k value above is safe */
6232 pOp = sqlite3VdbeGetOp(v, k - 1);
6233 assert( pOp->opcode!=OP_Column || pOp->p1!=pLevel->iTabCur );
6234 assert( pOp->opcode!=OP_Rowid || pOp->p1!=pLevel->iTabCur );
6235 assert( pOp->opcode!=OP_IfNullRow || pOp->p1!=pLevel->iTabCur );
drh299bf7c2018-06-11 17:35:02 +00006236#endif
drhcc04afd2013-08-22 02:56:28 +00006237 pOp = sqlite3VdbeGetOp(v, k);
drhf8556d02020-08-14 21:32:16 +00006238 pLastOp = pOp + (last - k);
drhe0cc2672021-04-05 22:42:15 +00006239 assert( pOp<=pLastOp );
drhf8556d02020-08-14 21:32:16 +00006240 do{
6241 if( pOp->p1!=pLevel->iTabCur ){
6242 /* no-op */
6243 }else if( pOp->opcode==OP_Column
drh092457b2017-12-29 15:04:49 +00006244#ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC
6245 || pOp->opcode==OP_Offset
6246#endif
6247 ){
drhee0ec8e2013-10-31 17:38:01 +00006248 int x = pOp->p2;
drh511717c2013-11-08 17:13:23 +00006249 assert( pIdx->pTable==pTab );
drh494317a2022-03-17 11:23:13 +00006250#ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC
6251 if( pOp->opcode==OP_Offset ){
6252 /* Do not need to translate the column number */
6253 }else
6254#endif
drhee0ec8e2013-10-31 17:38:01 +00006255 if( !HasRowid(pTab) ){
6256 Index *pPk = sqlite3PrimaryKeyIndex(pTab);
6257 x = pPk->aiColumn[x];
drh4b92f982015-09-29 17:20:14 +00006258 assert( x>=0 );
drh8e10d742019-10-18 17:42:47 +00006259 }else{
drhc5f808d2019-10-19 15:01:52 +00006260 testcase( x!=sqlite3StorageColumnToTable(pTab,x) );
drhb9bcf7c2019-10-19 13:29:10 +00006261 x = sqlite3StorageColumnToTable(pTab,x);
drhee0ec8e2013-10-31 17:38:01 +00006262 }
drhb9bcf7c2019-10-19 13:29:10 +00006263 x = sqlite3TableColumnToIndex(pIdx, x);
drh44156282013-10-23 22:23:03 +00006264 if( x>=0 ){
6265 pOp->p2 = x;
6266 pOp->p1 = pLevel->iIdxCur;
drh299bf7c2018-06-11 17:35:02 +00006267 OpcodeRewriteTrace(db, k, pOp);
drh75c493f2022-03-21 13:47:15 +00006268 }else{
6269 /* Unable to translate the table reference into an index
6270 ** reference. Verify that this is harmless - that the
6271 ** table being referenced really is open.
6272 */
drha753a3a2022-03-21 20:08:13 +00006273#ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC
drh75c493f2022-03-21 13:47:15 +00006274 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0
6275 || cursorIsOpen(v,pOp->p1,k)
drh67aa2312022-03-21 15:42:42 +00006276 || pOp->opcode==OP_Offset
drh75c493f2022-03-21 13:47:15 +00006277 );
drha753a3a2022-03-21 20:08:13 +00006278#else
6279 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0
6280 || cursorIsOpen(v,pOp->p1,k)
6281 );
6282#endif
drh9012bcb2004-12-19 00:11:35 +00006283 }
drhf0863fe2005-06-12 21:35:51 +00006284 }else if( pOp->opcode==OP_Rowid ){
drh9012bcb2004-12-19 00:11:35 +00006285 pOp->p1 = pLevel->iIdxCur;
drhf0863fe2005-06-12 21:35:51 +00006286 pOp->opcode = OP_IdxRowid;
drh299bf7c2018-06-11 17:35:02 +00006287 OpcodeRewriteTrace(db, k, pOp);
drh31d6fd52017-04-14 19:03:10 +00006288 }else if( pOp->opcode==OP_IfNullRow ){
6289 pOp->p1 = pLevel->iIdxCur;
drh299bf7c2018-06-11 17:35:02 +00006290 OpcodeRewriteTrace(db, k, pOp);
drh9012bcb2004-12-19 00:11:35 +00006291 }
drhf8556d02020-08-14 21:32:16 +00006292#ifdef SQLITE_DEBUG
6293 k++;
6294#endif
6295 }while( (++pOp)<pLastOp );
drh299bf7c2018-06-11 17:35:02 +00006296#ifdef SQLITE_DEBUG
6297 if( db->flags & SQLITE_VdbeAddopTrace ) printf("TRANSLATE complete\n");
6298#endif
drh6b563442001-11-07 16:48:26 +00006299 }
drh19a775c2000-06-05 18:54:46 +00006300 }
drh9012bcb2004-12-19 00:11:35 +00006301
drha76ac882022-04-08 19:20:12 +00006302 /* The "break" point is here, just past the end of the outer loop.
6303 ** Set it.
6304 */
6305 sqlite3VdbeResolveLabel(v, pWInfo->iBreak);
6306
drh9012bcb2004-12-19 00:11:35 +00006307 /* Final cleanup
6308 */
drhd784cc82021-04-15 12:56:44 +00006309 if( pWInfo->pExprMods ) whereUndoExprMods(pWInfo);
drhf12cde52010-04-08 17:28:00 +00006310 pParse->nQueryLoop = pWInfo->savedNQueryLoop;
6311 whereInfoFree(db, pWInfo);
drh75897232000-05-29 14:26:00 +00006312 return;
6313}