drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 1 | /* |
drh | b19a2bc | 2001-09-16 00:13:26 +0000 | [diff] [blame] | 2 | ** 2001 September 15 |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3 | ** |
drh | b19a2bc | 2001-09-16 00:13:26 +0000 | [diff] [blame] | 4 | ** The author disclaims copyright to this source code. In place of |
| 5 | ** a legal notice, here is a blessing: |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 6 | ** |
drh | b19a2bc | 2001-09-16 00:13:26 +0000 | [diff] [blame] | 7 | ** May you do good and not evil. |
| 8 | ** May you find forgiveness for yourself and forgive others. |
| 9 | ** May you share freely, never taking more than you give. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 10 | ** |
| 11 | ************************************************************************* |
| 12 | ** This module contains C code that generates VDBE code used to process |
drh | 909626d | 2008-05-30 14:58:37 +0000 | [diff] [blame] | 13 | ** the WHERE clause of SQL statements. This module is responsible for |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 14 | ** generating the code that loops through a table looking for applicable |
| 15 | ** rows. Indices are selected and used to speed the search when doing |
| 16 | ** so is applicable. Because this module is responsible for selecting |
| 17 | ** indices, you might also think of this module as the "query optimizer". |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 18 | ** |
danielk1977 | f51d1bd | 2009-07-31 06:14:51 +0000 | [diff] [blame] | 19 | ** $Id: where.c,v 1.411 2009/07/31 06:14:52 danielk1977 Exp $ |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 20 | */ |
| 21 | #include "sqliteInt.h" |
| 22 | |
| 23 | /* |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 24 | ** Trace output macros |
| 25 | */ |
| 26 | #if defined(SQLITE_TEST) || defined(SQLITE_DEBUG) |
mlcreech | 3a00f90 | 2008-03-04 17:45:01 +0000 | [diff] [blame] | 27 | int sqlite3WhereTrace = 0; |
drh | e8f52c5 | 2008-07-12 14:52:20 +0000 | [diff] [blame] | 28 | #endif |
drh | 85799a4 | 2009-04-07 13:48:11 +0000 | [diff] [blame] | 29 | #if defined(SQLITE_TEST) && defined(SQLITE_DEBUG) |
mlcreech | 3a00f90 | 2008-03-04 17:45:01 +0000 | [diff] [blame] | 30 | # define WHERETRACE(X) if(sqlite3WhereTrace) sqlite3DebugPrintf X |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 31 | #else |
drh | 4f0c587 | 2007-03-26 22:05:01 +0000 | [diff] [blame] | 32 | # define WHERETRACE(X) |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 33 | #endif |
| 34 | |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 35 | /* Forward reference |
| 36 | */ |
| 37 | typedef struct WhereClause WhereClause; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 38 | typedef struct WhereMaskSet WhereMaskSet; |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 39 | typedef struct WhereOrInfo WhereOrInfo; |
| 40 | typedef struct WhereAndInfo WhereAndInfo; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 41 | typedef struct WhereCost WhereCost; |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 42 | |
| 43 | /* |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 44 | ** The query generator uses an array of instances of this structure to |
| 45 | ** help it analyze the subexpressions of the WHERE clause. Each WHERE |
drh | 6149526 | 2009-04-22 15:32:59 +0000 | [diff] [blame] | 46 | ** clause subexpression is separated from the others by AND operators, |
| 47 | ** usually, or sometimes subexpressions separated by OR. |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 48 | ** |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 49 | ** All WhereTerms are collected into a single WhereClause structure. |
| 50 | ** The following identity holds: |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 51 | ** |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 52 | ** WhereTerm.pWC->a[WhereTerm.idx] == WhereTerm |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 53 | ** |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 54 | ** When a term is of the form: |
| 55 | ** |
| 56 | ** X <op> <expr> |
| 57 | ** |
| 58 | ** where X is a column name and <op> is one of certain operators, |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 59 | ** then WhereTerm.leftCursor and WhereTerm.u.leftColumn record the |
| 60 | ** cursor number and column number for X. WhereTerm.eOperator records |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 61 | ** the <op> using a bitmask encoding defined by WO_xxx below. The |
| 62 | ** use of a bitmask encoding for the operator allows us to search |
| 63 | ** quickly for terms that match any of several different operators. |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 64 | ** |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 65 | ** A WhereTerm might also be two or more subterms connected by OR: |
| 66 | ** |
| 67 | ** (t1.X <op> <expr>) OR (t1.Y <op> <expr>) OR .... |
| 68 | ** |
| 69 | ** In this second case, wtFlag as the TERM_ORINFO set and eOperator==WO_OR |
| 70 | ** and the WhereTerm.u.pOrInfo field points to auxiliary information that |
| 71 | ** is collected about the |
| 72 | ** |
| 73 | ** If a term in the WHERE clause does not match either of the two previous |
| 74 | ** categories, then eOperator==0. The WhereTerm.pExpr field is still set |
| 75 | ** to the original subexpression content and wtFlags is set up appropriately |
| 76 | ** but no other fields in the WhereTerm object are meaningful. |
| 77 | ** |
| 78 | ** When eOperator!=0, prereqRight and prereqAll record sets of cursor numbers, |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 79 | ** but they do so indirectly. A single WhereMaskSet structure translates |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 80 | ** cursor number into bits and the translated bit is stored in the prereq |
| 81 | ** fields. The translation is used in order to maximize the number of |
| 82 | ** bits that will fit in a Bitmask. The VDBE cursor numbers might be |
| 83 | ** spread out over the non-negative integers. For example, the cursor |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 84 | ** numbers might be 3, 8, 9, 10, 20, 23, 41, and 45. The WhereMaskSet |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 85 | ** translates these sparse cursor numbers into consecutive integers |
| 86 | ** beginning with 0 in order to make the best possible use of the available |
| 87 | ** bits in the Bitmask. So, in the example above, the cursor numbers |
| 88 | ** would be mapped into integers 0 through 7. |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 89 | ** |
| 90 | ** The number of terms in a join is limited by the number of bits |
| 91 | ** in prereqRight and prereqAll. The default is 64 bits, hence SQLite |
| 92 | ** is only able to process joins with 64 or fewer tables. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 93 | */ |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 94 | typedef struct WhereTerm WhereTerm; |
| 95 | struct WhereTerm { |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 96 | Expr *pExpr; /* Pointer to the subexpression that is this term */ |
drh | ec1724e | 2008-12-09 01:32:03 +0000 | [diff] [blame] | 97 | int iParent; /* Disable pWC->a[iParent] when this term disabled */ |
| 98 | int leftCursor; /* Cursor number of X in "X <op> <expr>" */ |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 99 | union { |
| 100 | int leftColumn; /* Column number of X in "X <op> <expr>" */ |
| 101 | WhereOrInfo *pOrInfo; /* Extra information if eOperator==WO_OR */ |
| 102 | WhereAndInfo *pAndInfo; /* Extra information if eOperator==WO_AND */ |
| 103 | } u; |
drh | b52076c | 2006-01-23 13:22:09 +0000 | [diff] [blame] | 104 | u16 eOperator; /* A WO_xx value describing <op> */ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 105 | u8 wtFlags; /* TERM_xxx bit flags. See below */ |
drh | 45b1ee4 | 2005-08-02 17:48:22 +0000 | [diff] [blame] | 106 | u8 nChild; /* Number of children that must disable us */ |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 107 | WhereClause *pWC; /* The clause this term is part of */ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 108 | Bitmask prereqRight; /* Bitmask of tables used by pExpr->pRight */ |
| 109 | Bitmask prereqAll; /* Bitmask of tables referenced by pExpr */ |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 110 | }; |
| 111 | |
| 112 | /* |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 113 | ** Allowed values of WhereTerm.wtFlags |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 114 | */ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 115 | #define TERM_DYNAMIC 0x01 /* Need to call sqlite3ExprDelete(db, pExpr) */ |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 116 | #define TERM_VIRTUAL 0x02 /* Added by the optimizer. Do not code */ |
| 117 | #define TERM_CODED 0x04 /* This term is already coded */ |
drh | 45b1ee4 | 2005-08-02 17:48:22 +0000 | [diff] [blame] | 118 | #define TERM_COPIED 0x08 /* Has a child */ |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 119 | #define TERM_ORINFO 0x10 /* Need to free the WhereTerm.u.pOrInfo object */ |
| 120 | #define TERM_ANDINFO 0x20 /* Need to free the WhereTerm.u.pAndInfo obj */ |
| 121 | #define TERM_OR_OK 0x40 /* Used during OR-clause processing */ |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 122 | |
| 123 | /* |
| 124 | ** An instance of the following structure holds all information about a |
| 125 | ** WHERE clause. Mostly this is a container for one or more WhereTerms. |
| 126 | */ |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 127 | struct WhereClause { |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 128 | Parse *pParse; /* The parser context */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 129 | WhereMaskSet *pMaskSet; /* Mapping of table cursor numbers to bitmasks */ |
danielk1977 | e672c8e | 2009-05-22 15:43:26 +0000 | [diff] [blame] | 130 | Bitmask vmask; /* Bitmask identifying virtual table cursors */ |
drh | 2943525 | 2008-12-28 18:35:08 +0000 | [diff] [blame] | 131 | u8 op; /* Split operator. TK_AND or TK_OR */ |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 132 | int nTerm; /* Number of terms */ |
| 133 | int nSlot; /* Number of entries in a[] */ |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 134 | WhereTerm *a; /* Each a[] describes a term of the WHERE cluase */ |
drh | 50d654d | 2009-06-03 01:24:54 +0000 | [diff] [blame] | 135 | #if defined(SQLITE_SMALL_STACK) |
| 136 | WhereTerm aStatic[1]; /* Initial static space for a[] */ |
| 137 | #else |
| 138 | WhereTerm aStatic[8]; /* Initial static space for a[] */ |
| 139 | #endif |
drh | e23399f | 2005-07-22 00:31:39 +0000 | [diff] [blame] | 140 | }; |
| 141 | |
| 142 | /* |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 143 | ** A WhereTerm with eOperator==WO_OR has its u.pOrInfo pointer set to |
| 144 | ** a dynamically allocated instance of the following structure. |
| 145 | */ |
| 146 | struct WhereOrInfo { |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 147 | WhereClause wc; /* Decomposition into subterms */ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 148 | Bitmask indexable; /* Bitmask of all indexable tables in the clause */ |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 149 | }; |
| 150 | |
| 151 | /* |
| 152 | ** A WhereTerm with eOperator==WO_AND has its u.pAndInfo pointer set to |
| 153 | ** a dynamically allocated instance of the following structure. |
| 154 | */ |
| 155 | struct WhereAndInfo { |
drh | 2943525 | 2008-12-28 18:35:08 +0000 | [diff] [blame] | 156 | WhereClause wc; /* The subexpression broken out */ |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 157 | }; |
| 158 | |
| 159 | /* |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 160 | ** An instance of the following structure keeps track of a mapping |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 161 | ** between VDBE cursor numbers and bits of the bitmasks in WhereTerm. |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 162 | ** |
| 163 | ** The VDBE cursor numbers are small integers contained in |
| 164 | ** SrcList_item.iCursor and Expr.iTable fields. For any given WHERE |
| 165 | ** clause, the cursor numbers might not begin with 0 and they might |
| 166 | ** contain gaps in the numbering sequence. But we want to make maximum |
| 167 | ** use of the bits in our bitmasks. This structure provides a mapping |
| 168 | ** from the sparse cursor numbers into consecutive integers beginning |
| 169 | ** with 0. |
| 170 | ** |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 171 | ** If WhereMaskSet.ix[A]==B it means that The A-th bit of a Bitmask |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 172 | ** corresponds VDBE cursor number B. The A-th bit of a bitmask is 1<<A. |
| 173 | ** |
| 174 | ** For example, if the WHERE clause expression used these VDBE |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 175 | ** cursors: 4, 5, 8, 29, 57, 73. Then the WhereMaskSet structure |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 176 | ** would map those cursor numbers into bits 0 through 5. |
| 177 | ** |
| 178 | ** Note that the mapping is not necessarily ordered. In the example |
| 179 | ** above, the mapping might go like this: 4->3, 5->1, 8->2, 29->0, |
| 180 | ** 57->5, 73->4. Or one of 719 other combinations might be used. It |
| 181 | ** does not really matter. What is important is that sparse cursor |
| 182 | ** numbers all get mapped into bit numbers that begin with 0 and contain |
| 183 | ** no gaps. |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 184 | */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 185 | struct WhereMaskSet { |
drh | 1398ad3 | 2005-01-19 23:24:50 +0000 | [diff] [blame] | 186 | int n; /* Number of assigned cursor values */ |
danielk1977 | 2343297 | 2008-11-17 16:42:00 +0000 | [diff] [blame] | 187 | int ix[BMS]; /* Cursor assigned to each bit */ |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 188 | }; |
| 189 | |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 190 | /* |
| 191 | ** A WhereCost object records a lookup strategy and the estimated |
| 192 | ** cost of pursuing that strategy. |
| 193 | */ |
| 194 | struct WhereCost { |
| 195 | WherePlan plan; /* The lookup strategy */ |
| 196 | double rCost; /* Overall cost of pursuing this search strategy */ |
| 197 | double nRow; /* Estimated number of output rows */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 198 | Bitmask used; /* Bitmask of cursors used by this plan */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 199 | }; |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 200 | |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 201 | /* |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 202 | ** Bitmasks for the operators that indices are able to exploit. An |
| 203 | ** OR-ed combination of these values can be used when searching for |
| 204 | ** terms in the where clause. |
| 205 | */ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 206 | #define WO_IN 0x001 |
| 207 | #define WO_EQ 0x002 |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 208 | #define WO_LT (WO_EQ<<(TK_LT-TK_EQ)) |
| 209 | #define WO_LE (WO_EQ<<(TK_LE-TK_EQ)) |
| 210 | #define WO_GT (WO_EQ<<(TK_GT-TK_EQ)) |
| 211 | #define WO_GE (WO_EQ<<(TK_GE-TK_EQ)) |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 212 | #define WO_MATCH 0x040 |
| 213 | #define WO_ISNULL 0x080 |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 214 | #define WO_OR 0x100 /* Two or more OR-connected terms */ |
| 215 | #define WO_AND 0x200 /* Two or more AND-connected terms */ |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 216 | |
drh | ec1724e | 2008-12-09 01:32:03 +0000 | [diff] [blame] | 217 | #define WO_ALL 0xfff /* Mask of all possible WO_* values */ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 218 | #define WO_SINGLE 0x0ff /* Mask of all non-compound WO_* values */ |
drh | ec1724e | 2008-12-09 01:32:03 +0000 | [diff] [blame] | 219 | |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 220 | /* |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 221 | ** Value for wsFlags returned by bestIndex() and stored in |
| 222 | ** WhereLevel.wsFlags. These flags determine which search |
| 223 | ** strategies are appropriate. |
drh | f2d315d | 2007-01-25 16:56:06 +0000 | [diff] [blame] | 224 | ** |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 225 | ** The least significant 12 bits is reserved as a mask for WO_ values above. |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 226 | ** The WhereLevel.wsFlags field is usually set to WO_IN|WO_EQ|WO_ISNULL. |
| 227 | ** But if the table is the right table of a left join, WhereLevel.wsFlags |
| 228 | ** is set to WO_IN|WO_EQ. The WhereLevel.wsFlags field can then be used as |
drh | f2d315d | 2007-01-25 16:56:06 +0000 | [diff] [blame] | 229 | ** the "op" parameter to findTerm when we are resolving equality constraints. |
| 230 | ** ISNULL constraints will then not be used on the right table of a left |
| 231 | ** join. Tickets #2177 and #2189. |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 232 | */ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 233 | #define WHERE_ROWID_EQ 0x00001000 /* rowid=EXPR or rowid IN (...) */ |
| 234 | #define WHERE_ROWID_RANGE 0x00002000 /* rowid<EXPR and/or rowid>EXPR */ |
drh | 46619d6 | 2009-04-24 14:51:42 +0000 | [diff] [blame] | 235 | #define WHERE_COLUMN_EQ 0x00010000 /* x=EXPR or x IN (...) or x IS NULL */ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 236 | #define WHERE_COLUMN_RANGE 0x00020000 /* x<EXPR and/or x>EXPR */ |
| 237 | #define WHERE_COLUMN_IN 0x00040000 /* x IN (...) */ |
drh | 46619d6 | 2009-04-24 14:51:42 +0000 | [diff] [blame] | 238 | #define WHERE_COLUMN_NULL 0x00080000 /* x IS NULL */ |
| 239 | #define WHERE_INDEXED 0x000f0000 /* Anything that uses an index */ |
| 240 | #define WHERE_IN_ABLE 0x000f1000 /* Able to support an IN operator */ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 241 | #define WHERE_TOP_LIMIT 0x00100000 /* x<EXPR or x<=EXPR constraint */ |
| 242 | #define WHERE_BTM_LIMIT 0x00200000 /* x>EXPR or x>=EXPR constraint */ |
| 243 | #define WHERE_IDX_ONLY 0x00800000 /* Use index only - omit table */ |
| 244 | #define WHERE_ORDERBY 0x01000000 /* Output will appear in correct order */ |
| 245 | #define WHERE_REVERSE 0x02000000 /* Scan in reverse order */ |
| 246 | #define WHERE_UNIQUE 0x04000000 /* Selects no more than one row */ |
| 247 | #define WHERE_VIRTUALTABLE 0x08000000 /* Use virtual-table processing */ |
| 248 | #define WHERE_MULTI_OR 0x10000000 /* OR using multiple indices */ |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 249 | |
| 250 | /* |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 251 | ** Initialize a preallocated WhereClause structure. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 252 | */ |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 253 | static void whereClauseInit( |
| 254 | WhereClause *pWC, /* The WhereClause to be initialized */ |
| 255 | Parse *pParse, /* The parsing context */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 256 | WhereMaskSet *pMaskSet /* Mapping from table cursor numbers to bitmasks */ |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 257 | ){ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 258 | pWC->pParse = pParse; |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 259 | pWC->pMaskSet = pMaskSet; |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 260 | pWC->nTerm = 0; |
drh | cad651e | 2007-04-20 12:22:01 +0000 | [diff] [blame] | 261 | pWC->nSlot = ArraySize(pWC->aStatic); |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 262 | pWC->a = pWC->aStatic; |
danielk1977 | e672c8e | 2009-05-22 15:43:26 +0000 | [diff] [blame] | 263 | pWC->vmask = 0; |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 264 | } |
| 265 | |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 266 | /* Forward reference */ |
| 267 | static void whereClauseClear(WhereClause*); |
| 268 | |
| 269 | /* |
| 270 | ** Deallocate all memory associated with a WhereOrInfo object. |
| 271 | */ |
| 272 | static void whereOrInfoDelete(sqlite3 *db, WhereOrInfo *p){ |
drh | 5bd98ae | 2009-01-07 18:24:03 +0000 | [diff] [blame] | 273 | whereClauseClear(&p->wc); |
| 274 | sqlite3DbFree(db, p); |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 275 | } |
| 276 | |
| 277 | /* |
| 278 | ** Deallocate all memory associated with a WhereAndInfo object. |
| 279 | */ |
| 280 | static void whereAndInfoDelete(sqlite3 *db, WhereAndInfo *p){ |
drh | 5bd98ae | 2009-01-07 18:24:03 +0000 | [diff] [blame] | 281 | whereClauseClear(&p->wc); |
| 282 | sqlite3DbFree(db, p); |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 283 | } |
| 284 | |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 285 | /* |
| 286 | ** Deallocate a WhereClause structure. The WhereClause structure |
| 287 | ** itself is not freed. This routine is the inverse of whereClauseInit(). |
| 288 | */ |
| 289 | static void whereClauseClear(WhereClause *pWC){ |
| 290 | int i; |
| 291 | WhereTerm *a; |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 292 | sqlite3 *db = pWC->pParse->db; |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 293 | for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 294 | if( a->wtFlags & TERM_DYNAMIC ){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 295 | sqlite3ExprDelete(db, a->pExpr); |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 296 | } |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 297 | if( a->wtFlags & TERM_ORINFO ){ |
| 298 | whereOrInfoDelete(db, a->u.pOrInfo); |
| 299 | }else if( a->wtFlags & TERM_ANDINFO ){ |
| 300 | whereAndInfoDelete(db, a->u.pAndInfo); |
| 301 | } |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 302 | } |
| 303 | if( pWC->a!=pWC->aStatic ){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 304 | sqlite3DbFree(db, pWC->a); |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 305 | } |
| 306 | } |
| 307 | |
| 308 | /* |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 309 | ** Add a single new WhereTerm entry to the WhereClause object pWC. |
| 310 | ** The new WhereTerm object is constructed from Expr p and with wtFlags. |
| 311 | ** The index in pWC->a[] of the new WhereTerm is returned on success. |
| 312 | ** 0 is returned if the new WhereTerm could not be added due to a memory |
| 313 | ** allocation error. The memory allocation failure will be recorded in |
| 314 | ** the db->mallocFailed flag so that higher-level functions can detect it. |
| 315 | ** |
| 316 | ** This routine will increase the size of the pWC->a[] array as necessary. |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 317 | ** |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 318 | ** If the wtFlags argument includes TERM_DYNAMIC, then responsibility |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 319 | ** for freeing the expression p is assumed by the WhereClause object pWC. |
| 320 | ** This is true even if this routine fails to allocate a new WhereTerm. |
drh | b63a53d | 2007-03-31 01:34:44 +0000 | [diff] [blame] | 321 | ** |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 322 | ** WARNING: This routine might reallocate the space used to store |
drh | 909626d | 2008-05-30 14:58:37 +0000 | [diff] [blame] | 323 | ** WhereTerms. All pointers to WhereTerms should be invalidated after |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 324 | ** calling this routine. Such pointers may be reinitialized by referencing |
| 325 | ** the pWC->a[] array. |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 326 | */ |
drh | ec1724e | 2008-12-09 01:32:03 +0000 | [diff] [blame] | 327 | static int whereClauseInsert(WhereClause *pWC, Expr *p, u8 wtFlags){ |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 328 | WhereTerm *pTerm; |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 329 | int idx; |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 330 | if( pWC->nTerm>=pWC->nSlot ){ |
| 331 | WhereTerm *pOld = pWC->a; |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 332 | sqlite3 *db = pWC->pParse->db; |
| 333 | pWC->a = sqlite3DbMallocRaw(db, sizeof(pWC->a[0])*pWC->nSlot*2 ); |
drh | b63a53d | 2007-03-31 01:34:44 +0000 | [diff] [blame] | 334 | if( pWC->a==0 ){ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 335 | if( wtFlags & TERM_DYNAMIC ){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 336 | sqlite3ExprDelete(db, p); |
drh | b63a53d | 2007-03-31 01:34:44 +0000 | [diff] [blame] | 337 | } |
drh | f998b73 | 2007-11-26 13:36:00 +0000 | [diff] [blame] | 338 | pWC->a = pOld; |
drh | b63a53d | 2007-03-31 01:34:44 +0000 | [diff] [blame] | 339 | return 0; |
| 340 | } |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 341 | memcpy(pWC->a, pOld, sizeof(pWC->a[0])*pWC->nTerm); |
| 342 | if( pOld!=pWC->aStatic ){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 343 | sqlite3DbFree(db, pOld); |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 344 | } |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 345 | pWC->nSlot = sqlite3DbMallocSize(db, pWC->a)/sizeof(pWC->a[0]); |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 346 | } |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 347 | pTerm = &pWC->a[idx = pWC->nTerm++]; |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 348 | pTerm->pExpr = p; |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 349 | pTerm->wtFlags = wtFlags; |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 350 | pTerm->pWC = pWC; |
drh | 45b1ee4 | 2005-08-02 17:48:22 +0000 | [diff] [blame] | 351 | pTerm->iParent = -1; |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 352 | return idx; |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 353 | } |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 354 | |
| 355 | /* |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 356 | ** This routine identifies subexpressions in the WHERE clause where |
drh | b6fb62d | 2005-09-20 08:47:20 +0000 | [diff] [blame] | 357 | ** each subexpression is separated by the AND operator or some other |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 358 | ** operator specified in the op parameter. The WhereClause structure |
| 359 | ** is filled with pointers to subexpressions. For example: |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 360 | ** |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 361 | ** WHERE a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22) |
| 362 | ** \________/ \_______________/ \________________/ |
| 363 | ** slot[0] slot[1] slot[2] |
| 364 | ** |
| 365 | ** The original WHERE clause in pExpr is unaltered. All this routine |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 366 | ** does is make slot[] entries point to substructure within pExpr. |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 367 | ** |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 368 | ** In the previous sentence and in the diagram, "slot[]" refers to |
drh | 902b9ee | 2008-12-05 17:17:07 +0000 | [diff] [blame] | 369 | ** the WhereClause.a[] array. The slot[] array grows as needed to contain |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 370 | ** all terms of the WHERE clause. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 371 | */ |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 372 | static void whereSplit(WhereClause *pWC, Expr *pExpr, int op){ |
drh | 2943525 | 2008-12-28 18:35:08 +0000 | [diff] [blame] | 373 | pWC->op = (u8)op; |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 374 | if( pExpr==0 ) return; |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 375 | if( pExpr->op!=op ){ |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 376 | whereClauseInsert(pWC, pExpr, 0); |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 377 | }else{ |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 378 | whereSplit(pWC, pExpr->pLeft, op); |
| 379 | whereSplit(pWC, pExpr->pRight, op); |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 380 | } |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 381 | } |
| 382 | |
| 383 | /* |
drh | 6149526 | 2009-04-22 15:32:59 +0000 | [diff] [blame] | 384 | ** Initialize an expression mask set (a WhereMaskSet object) |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 385 | */ |
| 386 | #define initMaskSet(P) memset(P, 0, sizeof(*P)) |
| 387 | |
| 388 | /* |
drh | 1398ad3 | 2005-01-19 23:24:50 +0000 | [diff] [blame] | 389 | ** Return the bitmask for the given cursor number. Return 0 if |
| 390 | ** iCursor is not in the set. |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 391 | */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 392 | static Bitmask getMask(WhereMaskSet *pMaskSet, int iCursor){ |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 393 | int i; |
drh | 3500ed6 | 2009-05-05 15:46:43 +0000 | [diff] [blame] | 394 | assert( pMaskSet->n<=sizeof(Bitmask)*8 ); |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 395 | for(i=0; i<pMaskSet->n; i++){ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 396 | if( pMaskSet->ix[i]==iCursor ){ |
| 397 | return ((Bitmask)1)<<i; |
| 398 | } |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 399 | } |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 400 | return 0; |
| 401 | } |
| 402 | |
| 403 | /* |
drh | 1398ad3 | 2005-01-19 23:24:50 +0000 | [diff] [blame] | 404 | ** Create a new mask for cursor iCursor. |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 405 | ** |
| 406 | ** There is one cursor per table in the FROM clause. The number of |
| 407 | ** tables in the FROM clause is limited by a test early in the |
drh | b6fb62d | 2005-09-20 08:47:20 +0000 | [diff] [blame] | 408 | ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 409 | ** array will never overflow. |
drh | 1398ad3 | 2005-01-19 23:24:50 +0000 | [diff] [blame] | 410 | */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 411 | static void createMask(WhereMaskSet *pMaskSet, int iCursor){ |
drh | cad651e | 2007-04-20 12:22:01 +0000 | [diff] [blame] | 412 | assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 413 | pMaskSet->ix[pMaskSet->n++] = iCursor; |
drh | 1398ad3 | 2005-01-19 23:24:50 +0000 | [diff] [blame] | 414 | } |
| 415 | |
| 416 | /* |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 417 | ** This routine walks (recursively) an expression tree and generates |
| 418 | ** a bitmask indicating which tables are used in that expression |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 419 | ** tree. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 420 | ** |
| 421 | ** In order for this routine to work, the calling function must have |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 422 | ** previously invoked sqlite3ResolveExprNames() on the expression. See |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 423 | ** the header comment on that routine for additional information. |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 424 | ** The sqlite3ResolveExprNames() routines looks for column names and |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 425 | ** sets their opcodes to TK_COLUMN and their Expr.iTable fields to |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 426 | ** the VDBE cursor number of the table. This routine just has to |
| 427 | ** translate the cursor numbers into bitmask values and OR all |
| 428 | ** the bitmasks together. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 429 | */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 430 | static Bitmask exprListTableUsage(WhereMaskSet*, ExprList*); |
| 431 | static Bitmask exprSelectTableUsage(WhereMaskSet*, Select*); |
| 432 | static Bitmask exprTableUsage(WhereMaskSet *pMaskSet, Expr *p){ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 433 | Bitmask mask = 0; |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 434 | if( p==0 ) return 0; |
drh | 967e8b7 | 2000-06-21 13:59:10 +0000 | [diff] [blame] | 435 | if( p->op==TK_COLUMN ){ |
drh | 8feb4b1 | 2004-07-19 02:12:14 +0000 | [diff] [blame] | 436 | mask = getMask(pMaskSet, p->iTable); |
drh | 8feb4b1 | 2004-07-19 02:12:14 +0000 | [diff] [blame] | 437 | return mask; |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 438 | } |
danielk1977 | b3bce66 | 2005-01-29 08:32:43 +0000 | [diff] [blame] | 439 | mask = exprTableUsage(pMaskSet, p->pRight); |
| 440 | mask |= exprTableUsage(pMaskSet, p->pLeft); |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 441 | if( ExprHasProperty(p, EP_xIsSelect) ){ |
| 442 | mask |= exprSelectTableUsage(pMaskSet, p->x.pSelect); |
| 443 | }else{ |
| 444 | mask |= exprListTableUsage(pMaskSet, p->x.pList); |
| 445 | } |
danielk1977 | b3bce66 | 2005-01-29 08:32:43 +0000 | [diff] [blame] | 446 | return mask; |
| 447 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 448 | static Bitmask exprListTableUsage(WhereMaskSet *pMaskSet, ExprList *pList){ |
danielk1977 | b3bce66 | 2005-01-29 08:32:43 +0000 | [diff] [blame] | 449 | int i; |
| 450 | Bitmask mask = 0; |
| 451 | if( pList ){ |
| 452 | for(i=0; i<pList->nExpr; i++){ |
| 453 | mask |= exprTableUsage(pMaskSet, pList->a[i].pExpr); |
drh | dd57912 | 2002-04-02 01:58:57 +0000 | [diff] [blame] | 454 | } |
| 455 | } |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 456 | return mask; |
| 457 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 458 | static Bitmask exprSelectTableUsage(WhereMaskSet *pMaskSet, Select *pS){ |
drh | a430ae8 | 2007-09-12 15:41:01 +0000 | [diff] [blame] | 459 | Bitmask mask = 0; |
| 460 | while( pS ){ |
| 461 | mask |= exprListTableUsage(pMaskSet, pS->pEList); |
drh | f5b1138 | 2005-09-17 13:07:13 +0000 | [diff] [blame] | 462 | mask |= exprListTableUsage(pMaskSet, pS->pGroupBy); |
| 463 | mask |= exprListTableUsage(pMaskSet, pS->pOrderBy); |
| 464 | mask |= exprTableUsage(pMaskSet, pS->pWhere); |
| 465 | mask |= exprTableUsage(pMaskSet, pS->pHaving); |
drh | a430ae8 | 2007-09-12 15:41:01 +0000 | [diff] [blame] | 466 | pS = pS->pPrior; |
drh | f5b1138 | 2005-09-17 13:07:13 +0000 | [diff] [blame] | 467 | } |
| 468 | return mask; |
| 469 | } |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 470 | |
| 471 | /* |
drh | 487ab3c | 2001-11-08 00:45:21 +0000 | [diff] [blame] | 472 | ** Return TRUE if the given operator is one of the operators that is |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 473 | ** allowed for an indexable WHERE clause term. The allowed operators are |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 474 | ** "=", "<", ">", "<=", ">=", and "IN". |
drh | 487ab3c | 2001-11-08 00:45:21 +0000 | [diff] [blame] | 475 | */ |
| 476 | static int allowedOp(int op){ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 477 | assert( TK_GT>TK_EQ && TK_GT<TK_GE ); |
| 478 | assert( TK_LT>TK_EQ && TK_LT<TK_GE ); |
| 479 | assert( TK_LE>TK_EQ && TK_LE<TK_GE ); |
| 480 | assert( TK_GE==TK_EQ+4 ); |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 481 | return op==TK_IN || (op>=TK_EQ && op<=TK_GE) || op==TK_ISNULL; |
drh | 487ab3c | 2001-11-08 00:45:21 +0000 | [diff] [blame] | 482 | } |
| 483 | |
| 484 | /* |
drh | 902b9ee | 2008-12-05 17:17:07 +0000 | [diff] [blame] | 485 | ** Swap two objects of type TYPE. |
drh | 193bd77 | 2004-07-20 18:23:14 +0000 | [diff] [blame] | 486 | */ |
| 487 | #define SWAP(TYPE,A,B) {TYPE t=A; A=B; B=t;} |
| 488 | |
| 489 | /* |
drh | 909626d | 2008-05-30 14:58:37 +0000 | [diff] [blame] | 490 | ** Commute a comparison operator. Expressions of the form "X op Y" |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 491 | ** are converted into "Y op X". |
danielk1977 | eb5453d | 2007-07-30 14:40:48 +0000 | [diff] [blame] | 492 | ** |
| 493 | ** If a collation sequence is associated with either the left or right |
| 494 | ** side of the comparison, it remains associated with the same side after |
| 495 | ** the commutation. So "Y collate NOCASE op X" becomes |
| 496 | ** "X collate NOCASE op Y". This is because any collation sequence on |
| 497 | ** the left hand side of a comparison overrides any collation sequence |
| 498 | ** attached to the right. For the same reason the EP_ExpCollate flag |
| 499 | ** is not commuted. |
drh | 193bd77 | 2004-07-20 18:23:14 +0000 | [diff] [blame] | 500 | */ |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 501 | static void exprCommute(Parse *pParse, Expr *pExpr){ |
danielk1977 | eb5453d | 2007-07-30 14:40:48 +0000 | [diff] [blame] | 502 | u16 expRight = (pExpr->pRight->flags & EP_ExpCollate); |
| 503 | u16 expLeft = (pExpr->pLeft->flags & EP_ExpCollate); |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 504 | assert( allowedOp(pExpr->op) && pExpr->op!=TK_IN ); |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 505 | pExpr->pRight->pColl = sqlite3ExprCollSeq(pParse, pExpr->pRight); |
| 506 | pExpr->pLeft->pColl = sqlite3ExprCollSeq(pParse, pExpr->pLeft); |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 507 | SWAP(CollSeq*,pExpr->pRight->pColl,pExpr->pLeft->pColl); |
danielk1977 | eb5453d | 2007-07-30 14:40:48 +0000 | [diff] [blame] | 508 | pExpr->pRight->flags = (pExpr->pRight->flags & ~EP_ExpCollate) | expLeft; |
| 509 | pExpr->pLeft->flags = (pExpr->pLeft->flags & ~EP_ExpCollate) | expRight; |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 510 | SWAP(Expr*,pExpr->pRight,pExpr->pLeft); |
| 511 | if( pExpr->op>=TK_GT ){ |
| 512 | assert( TK_LT==TK_GT+2 ); |
| 513 | assert( TK_GE==TK_LE+2 ); |
| 514 | assert( TK_GT>TK_EQ ); |
| 515 | assert( TK_GT<TK_LE ); |
| 516 | assert( pExpr->op>=TK_GT && pExpr->op<=TK_GE ); |
| 517 | pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT; |
drh | 193bd77 | 2004-07-20 18:23:14 +0000 | [diff] [blame] | 518 | } |
drh | 193bd77 | 2004-07-20 18:23:14 +0000 | [diff] [blame] | 519 | } |
| 520 | |
| 521 | /* |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 522 | ** Translate from TK_xx operator to WO_xx bitmask. |
| 523 | */ |
drh | ec1724e | 2008-12-09 01:32:03 +0000 | [diff] [blame] | 524 | static u16 operatorMask(int op){ |
| 525 | u16 c; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 526 | assert( allowedOp(op) ); |
| 527 | if( op==TK_IN ){ |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 528 | c = WO_IN; |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 529 | }else if( op==TK_ISNULL ){ |
| 530 | c = WO_ISNULL; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 531 | }else{ |
drh | ec1724e | 2008-12-09 01:32:03 +0000 | [diff] [blame] | 532 | assert( (WO_EQ<<(op-TK_EQ)) < 0x7fff ); |
| 533 | c = (u16)(WO_EQ<<(op-TK_EQ)); |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 534 | } |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 535 | assert( op!=TK_ISNULL || c==WO_ISNULL ); |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 536 | assert( op!=TK_IN || c==WO_IN ); |
| 537 | assert( op!=TK_EQ || c==WO_EQ ); |
| 538 | assert( op!=TK_LT || c==WO_LT ); |
| 539 | assert( op!=TK_LE || c==WO_LE ); |
| 540 | assert( op!=TK_GT || c==WO_GT ); |
| 541 | assert( op!=TK_GE || c==WO_GE ); |
| 542 | return c; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 543 | } |
| 544 | |
| 545 | /* |
| 546 | ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" |
| 547 | ** where X is a reference to the iColumn of table iCur and <op> is one of |
| 548 | ** the WO_xx operator codes specified by the op parameter. |
| 549 | ** Return a pointer to the term. Return 0 if not found. |
| 550 | */ |
| 551 | static WhereTerm *findTerm( |
| 552 | WhereClause *pWC, /* The WHERE clause to be searched */ |
| 553 | int iCur, /* Cursor number of LHS */ |
| 554 | int iColumn, /* Column number of LHS */ |
| 555 | Bitmask notReady, /* RHS must not overlap with this mask */ |
drh | ec1724e | 2008-12-09 01:32:03 +0000 | [diff] [blame] | 556 | u32 op, /* Mask of WO_xx values describing operator */ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 557 | Index *pIdx /* Must be compatible with this index, if not NULL */ |
| 558 | ){ |
| 559 | WhereTerm *pTerm; |
| 560 | int k; |
drh | 22c2403 | 2008-07-09 13:28:53 +0000 | [diff] [blame] | 561 | assert( iCur>=0 ); |
drh | ec1724e | 2008-12-09 01:32:03 +0000 | [diff] [blame] | 562 | op &= WO_ALL; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 563 | for(pTerm=pWC->a, k=pWC->nTerm; k; k--, pTerm++){ |
| 564 | if( pTerm->leftCursor==iCur |
| 565 | && (pTerm->prereqRight & notReady)==0 |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 566 | && pTerm->u.leftColumn==iColumn |
drh | b52076c | 2006-01-23 13:22:09 +0000 | [diff] [blame] | 567 | && (pTerm->eOperator & op)!=0 |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 568 | ){ |
drh | 22c2403 | 2008-07-09 13:28:53 +0000 | [diff] [blame] | 569 | if( pIdx && pTerm->eOperator!=WO_ISNULL ){ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 570 | Expr *pX = pTerm->pExpr; |
| 571 | CollSeq *pColl; |
| 572 | char idxaff; |
danielk1977 | f011300 | 2006-01-24 12:09:17 +0000 | [diff] [blame] | 573 | int j; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 574 | Parse *pParse = pWC->pParse; |
| 575 | |
| 576 | idxaff = pIdx->pTable->aCol[iColumn].affinity; |
| 577 | if( !sqlite3IndexAffinityOk(pX, idxaff) ) continue; |
danielk1977 | bcbb04e | 2007-05-29 12:11:29 +0000 | [diff] [blame] | 578 | |
| 579 | /* Figure out the collation sequence required from an index for |
| 580 | ** it to be useful for optimising expression pX. Store this |
| 581 | ** value in variable pColl. |
| 582 | */ |
| 583 | assert(pX->pLeft); |
| 584 | pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); |
danielk1977 | 9357416 | 2008-12-30 15:26:29 +0000 | [diff] [blame] | 585 | assert(pColl || pParse->nErr); |
danielk1977 | bcbb04e | 2007-05-29 12:11:29 +0000 | [diff] [blame] | 586 | |
drh | 22c2403 | 2008-07-09 13:28:53 +0000 | [diff] [blame] | 587 | for(j=0; pIdx->aiColumn[j]!=iColumn; j++){ |
drh | 34004ce | 2008-07-11 16:15:17 +0000 | [diff] [blame] | 588 | if( NEVER(j>=pIdx->nColumn) ) return 0; |
drh | 22c2403 | 2008-07-09 13:28:53 +0000 | [diff] [blame] | 589 | } |
danielk1977 | 9357416 | 2008-12-30 15:26:29 +0000 | [diff] [blame] | 590 | if( pColl && sqlite3StrICmp(pColl->zName, pIdx->azColl[j]) ) continue; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 591 | } |
| 592 | return pTerm; |
| 593 | } |
| 594 | } |
| 595 | return 0; |
| 596 | } |
| 597 | |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 598 | /* Forward reference */ |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 599 | static void exprAnalyze(SrcList*, WhereClause*, int); |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 600 | |
| 601 | /* |
| 602 | ** Call exprAnalyze on all terms in a WHERE clause. |
| 603 | ** |
| 604 | ** |
| 605 | */ |
| 606 | static void exprAnalyzeAll( |
| 607 | SrcList *pTabList, /* the FROM clause */ |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 608 | WhereClause *pWC /* the WHERE clause to be analyzed */ |
| 609 | ){ |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 610 | int i; |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 611 | for(i=pWC->nTerm-1; i>=0; i--){ |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 612 | exprAnalyze(pTabList, pWC, i); |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 613 | } |
| 614 | } |
| 615 | |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 616 | #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION |
| 617 | /* |
| 618 | ** Check to see if the given expression is a LIKE or GLOB operator that |
| 619 | ** can be optimized using inequality constraints. Return TRUE if it is |
| 620 | ** so and false if not. |
| 621 | ** |
| 622 | ** In order for the operator to be optimizible, the RHS must be a string |
| 623 | ** literal that does not begin with a wildcard. |
| 624 | */ |
| 625 | static int isLikeOrGlob( |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 626 | Parse *pParse, /* Parsing and code generating context */ |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 627 | Expr *pExpr, /* Test this expression */ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 628 | Expr **ppPrefix, /* Pointer to TK_STRING expression with pattern prefix */ |
drh | 9f504ea | 2008-02-23 21:55:39 +0000 | [diff] [blame] | 629 | int *pisComplete, /* True if the only wildcard is % in the last character */ |
| 630 | int *pnoCase /* True if uppercase is equivalent to lowercase */ |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 631 | ){ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 632 | const char *z = 0; /* String on RHS of LIKE operator */ |
drh | 5bd98ae | 2009-01-07 18:24:03 +0000 | [diff] [blame] | 633 | Expr *pRight, *pLeft; /* Right and left size of LIKE operator */ |
| 634 | ExprList *pList; /* List of operands to the LIKE operator */ |
| 635 | int c; /* One character in z[] */ |
| 636 | int cnt; /* Number of non-wildcard prefix characters */ |
| 637 | char wc[3]; /* Wildcard characters */ |
| 638 | CollSeq *pColl; /* Collating sequence for LHS */ |
| 639 | sqlite3 *db = pParse->db; /* Database connection */ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 640 | sqlite3_value *pVal = 0; |
| 641 | int op; /* Opcode of pRight */ |
drh | d64fe2f | 2005-08-28 17:00:23 +0000 | [diff] [blame] | 642 | |
drh | 9f504ea | 2008-02-23 21:55:39 +0000 | [diff] [blame] | 643 | if( !sqlite3IsLikeFunction(db, pExpr, pnoCase, wc) ){ |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 644 | return 0; |
| 645 | } |
drh | 9f504ea | 2008-02-23 21:55:39 +0000 | [diff] [blame] | 646 | #ifdef SQLITE_EBCDIC |
| 647 | if( *pnoCase ) return 0; |
| 648 | #endif |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 649 | pList = pExpr->x.pList; |
drh | 55ef4d9 | 2005-08-14 01:20:37 +0000 | [diff] [blame] | 650 | pLeft = pList->a[1].pExpr; |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 651 | if( pLeft->op!=TK_COLUMN ){ |
| 652 | return 0; |
| 653 | } |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 654 | pColl = sqlite3ExprCollSeq(pParse, pLeft); |
drh | 01495b9 | 2008-01-23 12:52:40 +0000 | [diff] [blame] | 655 | assert( pColl!=0 || pLeft->iColumn==-1 ); |
drh | c4ac22e | 2009-06-07 23:45:10 +0000 | [diff] [blame] | 656 | if( pColl==0 ) return 0; |
drh | 9f504ea | 2008-02-23 21:55:39 +0000 | [diff] [blame] | 657 | if( (pColl->type!=SQLITE_COLL_BINARY || *pnoCase) && |
| 658 | (pColl->type!=SQLITE_COLL_NOCASE || !*pnoCase) ){ |
drh | d64fe2f | 2005-08-28 17:00:23 +0000 | [diff] [blame] | 659 | return 0; |
| 660 | } |
drh | c4ac22e | 2009-06-07 23:45:10 +0000 | [diff] [blame] | 661 | if( sqlite3ExprAffinity(pLeft)!=SQLITE_AFF_TEXT ) return 0; |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 662 | |
| 663 | pRight = pList->a[0].pExpr; |
| 664 | op = pRight->op; |
| 665 | if( op==TK_REGISTER ){ |
| 666 | op = pRight->op2; |
| 667 | } |
| 668 | if( op==TK_VARIABLE ){ |
| 669 | Vdbe *pReprepare = pParse->pReprepare; |
| 670 | pVal = sqlite3VdbeGetValue(pReprepare, pRight->iColumn, SQLITE_AFF_NONE); |
| 671 | if( pVal && sqlite3_value_type(pVal)==SQLITE_TEXT ){ |
| 672 | z = (char *)sqlite3_value_text(pVal); |
| 673 | } |
dan | 1d2ce4f | 2009-10-19 18:11:09 +0000 | [diff] [blame] | 674 | sqlite3VdbeSetVarmask(pParse->pVdbe, pRight->iColumn); |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 675 | assert( pRight->op==TK_VARIABLE || pRight->op==TK_REGISTER ); |
| 676 | }else if( op==TK_STRING ){ |
| 677 | z = pRight->u.zToken; |
| 678 | } |
| 679 | if( z ){ |
shane | 8509570 | 2009-06-15 16:27:08 +0000 | [diff] [blame] | 680 | cnt = 0; |
drh | b7916a7 | 2009-05-27 10:31:29 +0000 | [diff] [blame] | 681 | while( (c=z[cnt])!=0 && c!=wc[0] && c!=wc[1] && c!=wc[2] ){ |
drh | 24fb627 | 2009-05-01 21:13:36 +0000 | [diff] [blame] | 682 | cnt++; |
| 683 | } |
shane | 8509570 | 2009-06-15 16:27:08 +0000 | [diff] [blame] | 684 | if( cnt!=0 && c!=0 && 255!=(u8)z[cnt-1] ){ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 685 | Expr *pPrefix; |
shane | 8509570 | 2009-06-15 16:27:08 +0000 | [diff] [blame] | 686 | *pisComplete = z[cnt]==wc[0] && z[cnt+1]==0; |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 687 | pPrefix = sqlite3Expr(db, TK_STRING, z); |
| 688 | if( pPrefix ) pPrefix->u.zToken[cnt] = 0; |
| 689 | *ppPrefix = pPrefix; |
| 690 | if( op==TK_VARIABLE ){ |
| 691 | Vdbe *v = pParse->pVdbe; |
dan | 1d2ce4f | 2009-10-19 18:11:09 +0000 | [diff] [blame] | 692 | sqlite3VdbeSetVarmask(v, pRight->iColumn); |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 693 | if( *pisComplete && pRight->u.zToken[1] ){ |
| 694 | /* If the rhs of the LIKE expression is a variable, and the current |
| 695 | ** value of the variable means there is no need to invoke the LIKE |
| 696 | ** function, then no OP_Variable will be added to the program. |
| 697 | ** This causes problems for the sqlite3_bind_parameter_name() |
drh | bec451f | 2009-10-17 13:13:02 +0000 | [diff] [blame] | 698 | ** API. To workaround them, add a dummy OP_Variable here. |
| 699 | */ |
| 700 | int r1 = sqlite3GetTempReg(pParse); |
| 701 | sqlite3ExprCodeTarget(pParse, pRight, r1); |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 702 | sqlite3VdbeChangeP3(v, sqlite3VdbeCurrentAddr(v)-1, 0); |
drh | bec451f | 2009-10-17 13:13:02 +0000 | [diff] [blame] | 703 | sqlite3ReleaseTempReg(pParse, r1); |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 704 | } |
| 705 | } |
| 706 | }else{ |
| 707 | z = 0; |
shane | 8509570 | 2009-06-15 16:27:08 +0000 | [diff] [blame] | 708 | } |
drh | f998b73 | 2007-11-26 13:36:00 +0000 | [diff] [blame] | 709 | } |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 710 | |
| 711 | sqlite3ValueFree(pVal); |
| 712 | return (z!=0); |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 713 | } |
| 714 | #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */ |
| 715 | |
drh | edb193b | 2006-06-27 13:20:21 +0000 | [diff] [blame] | 716 | |
| 717 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 718 | /* |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 719 | ** Check to see if the given expression is of the form |
| 720 | ** |
| 721 | ** column MATCH expr |
| 722 | ** |
| 723 | ** If it is then return TRUE. If not, return FALSE. |
| 724 | */ |
| 725 | static int isMatchOfColumn( |
| 726 | Expr *pExpr /* Test this expression */ |
| 727 | ){ |
| 728 | ExprList *pList; |
| 729 | |
| 730 | if( pExpr->op!=TK_FUNCTION ){ |
| 731 | return 0; |
| 732 | } |
drh | 33e619f | 2009-05-28 01:00:55 +0000 | [diff] [blame] | 733 | if( sqlite3StrICmp(pExpr->u.zToken,"match")!=0 ){ |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 734 | return 0; |
| 735 | } |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 736 | pList = pExpr->x.pList; |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 737 | if( pList->nExpr!=2 ){ |
| 738 | return 0; |
| 739 | } |
| 740 | if( pList->a[1].pExpr->op != TK_COLUMN ){ |
| 741 | return 0; |
| 742 | } |
| 743 | return 1; |
| 744 | } |
drh | edb193b | 2006-06-27 13:20:21 +0000 | [diff] [blame] | 745 | #endif /* SQLITE_OMIT_VIRTUALTABLE */ |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 746 | |
| 747 | /* |
drh | 54a167d | 2005-11-26 14:08:07 +0000 | [diff] [blame] | 748 | ** If the pBase expression originated in the ON or USING clause of |
| 749 | ** a join, then transfer the appropriate markings over to derived. |
| 750 | */ |
| 751 | static void transferJoinMarkings(Expr *pDerived, Expr *pBase){ |
| 752 | pDerived->flags |= pBase->flags & EP_FromJoin; |
| 753 | pDerived->iRightJoinTable = pBase->iRightJoinTable; |
| 754 | } |
| 755 | |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 756 | #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY) |
| 757 | /* |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 758 | ** Analyze a term that consists of two or more OR-connected |
| 759 | ** subterms. So in: |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 760 | ** |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 761 | ** ... WHERE (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13) |
| 762 | ** ^^^^^^^^^^^^^^^^^^^^ |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 763 | ** |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 764 | ** This routine analyzes terms such as the middle term in the above example. |
| 765 | ** A WhereOrTerm object is computed and attached to the term under |
| 766 | ** analysis, regardless of the outcome of the analysis. Hence: |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 767 | ** |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 768 | ** WhereTerm.wtFlags |= TERM_ORINFO |
| 769 | ** WhereTerm.u.pOrInfo = a dynamically allocated WhereOrTerm object |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 770 | ** |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 771 | ** The term being analyzed must have two or more of OR-connected subterms. |
danielk1977 | fdc4019 | 2008-12-29 18:33:32 +0000 | [diff] [blame] | 772 | ** A single subterm might be a set of AND-connected sub-subterms. |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 773 | ** Examples of terms under analysis: |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 774 | ** |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 775 | ** (A) t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5 |
| 776 | ** (B) x=expr1 OR expr2=x OR x=expr3 |
| 777 | ** (C) t1.x=t2.y OR (t1.x=t2.z AND t1.y=15) |
| 778 | ** (D) x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*') |
| 779 | ** (E) (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6) |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 780 | ** |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 781 | ** CASE 1: |
| 782 | ** |
| 783 | ** If all subterms are of the form T.C=expr for some single column of C |
| 784 | ** a single table T (as shown in example B above) then create a new virtual |
| 785 | ** term that is an equivalent IN expression. In other words, if the term |
| 786 | ** being analyzed is: |
| 787 | ** |
| 788 | ** x = expr1 OR expr2 = x OR x = expr3 |
| 789 | ** |
| 790 | ** then create a new virtual term like this: |
| 791 | ** |
| 792 | ** x IN (expr1,expr2,expr3) |
| 793 | ** |
| 794 | ** CASE 2: |
| 795 | ** |
| 796 | ** If all subterms are indexable by a single table T, then set |
| 797 | ** |
| 798 | ** WhereTerm.eOperator = WO_OR |
| 799 | ** WhereTerm.u.pOrInfo->indexable |= the cursor number for table T |
| 800 | ** |
| 801 | ** A subterm is "indexable" if it is of the form |
| 802 | ** "T.C <op> <expr>" where C is any column of table T and |
| 803 | ** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN". |
| 804 | ** A subterm is also indexable if it is an AND of two or more |
| 805 | ** subsubterms at least one of which is indexable. Indexable AND |
| 806 | ** subterms have their eOperator set to WO_AND and they have |
| 807 | ** u.pAndInfo set to a dynamically allocated WhereAndTerm object. |
| 808 | ** |
| 809 | ** From another point of view, "indexable" means that the subterm could |
| 810 | ** potentially be used with an index if an appropriate index exists. |
| 811 | ** This analysis does not consider whether or not the index exists; that |
| 812 | ** is something the bestIndex() routine will determine. This analysis |
| 813 | ** only looks at whether subterms appropriate for indexing exist. |
| 814 | ** |
| 815 | ** All examples A through E above all satisfy case 2. But if a term |
| 816 | ** also statisfies case 1 (such as B) we know that the optimizer will |
| 817 | ** always prefer case 1, so in that case we pretend that case 2 is not |
| 818 | ** satisfied. |
| 819 | ** |
| 820 | ** It might be the case that multiple tables are indexable. For example, |
| 821 | ** (E) above is indexable on tables P, Q, and R. |
| 822 | ** |
| 823 | ** Terms that satisfy case 2 are candidates for lookup by using |
| 824 | ** separate indices to find rowids for each subterm and composing |
| 825 | ** the union of all rowids using a RowSet object. This is similar |
| 826 | ** to "bitmap indices" in other database engines. |
| 827 | ** |
| 828 | ** OTHERWISE: |
| 829 | ** |
| 830 | ** If neither case 1 nor case 2 apply, then leave the eOperator set to |
| 831 | ** zero. This term is not useful for search. |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 832 | */ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 833 | static void exprAnalyzeOrTerm( |
| 834 | SrcList *pSrc, /* the FROM clause */ |
| 835 | WhereClause *pWC, /* the complete WHERE clause */ |
| 836 | int idxTerm /* Index of the OR-term to be analyzed */ |
| 837 | ){ |
| 838 | Parse *pParse = pWC->pParse; /* Parser context */ |
| 839 | sqlite3 *db = pParse->db; /* Database connection */ |
| 840 | WhereTerm *pTerm = &pWC->a[idxTerm]; /* The term to be analyzed */ |
| 841 | Expr *pExpr = pTerm->pExpr; /* The expression of the term */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 842 | WhereMaskSet *pMaskSet = pWC->pMaskSet; /* Table use masks */ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 843 | int i; /* Loop counters */ |
| 844 | WhereClause *pOrWc; /* Breakup of pTerm into subterms */ |
| 845 | WhereTerm *pOrTerm; /* A Sub-term within the pOrWc */ |
| 846 | WhereOrInfo *pOrInfo; /* Additional information associated with pTerm */ |
| 847 | Bitmask chngToIN; /* Tables that might satisfy case 1 */ |
| 848 | Bitmask indexable; /* Tables that are indexable, satisfying case 2 */ |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 849 | |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 850 | /* |
| 851 | ** Break the OR clause into its separate subterms. The subterms are |
| 852 | ** stored in a WhereClause structure containing within the WhereOrInfo |
| 853 | ** object that is attached to the original OR clause term. |
| 854 | */ |
| 855 | assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 ); |
| 856 | assert( pExpr->op==TK_OR ); |
drh | 954701a | 2008-12-29 23:45:07 +0000 | [diff] [blame] | 857 | pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo)); |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 858 | if( pOrInfo==0 ) return; |
| 859 | pTerm->wtFlags |= TERM_ORINFO; |
| 860 | pOrWc = &pOrInfo->wc; |
| 861 | whereClauseInit(pOrWc, pWC->pParse, pMaskSet); |
| 862 | whereSplit(pOrWc, pExpr, TK_OR); |
| 863 | exprAnalyzeAll(pSrc, pOrWc); |
| 864 | if( db->mallocFailed ) return; |
| 865 | assert( pOrWc->nTerm>=2 ); |
| 866 | |
| 867 | /* |
| 868 | ** Compute the set of tables that might satisfy cases 1 or 2. |
| 869 | */ |
danielk1977 | e672c8e | 2009-05-22 15:43:26 +0000 | [diff] [blame] | 870 | indexable = ~(Bitmask)0; |
| 871 | chngToIN = ~(pWC->vmask); |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 872 | for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){ |
| 873 | if( (pOrTerm->eOperator & WO_SINGLE)==0 ){ |
drh | 2943525 | 2008-12-28 18:35:08 +0000 | [diff] [blame] | 874 | WhereAndInfo *pAndInfo; |
| 875 | assert( pOrTerm->eOperator==0 ); |
| 876 | assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 ); |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 877 | chngToIN = 0; |
drh | 2943525 | 2008-12-28 18:35:08 +0000 | [diff] [blame] | 878 | pAndInfo = sqlite3DbMallocRaw(db, sizeof(*pAndInfo)); |
| 879 | if( pAndInfo ){ |
| 880 | WhereClause *pAndWC; |
| 881 | WhereTerm *pAndTerm; |
| 882 | int j; |
| 883 | Bitmask b = 0; |
| 884 | pOrTerm->u.pAndInfo = pAndInfo; |
| 885 | pOrTerm->wtFlags |= TERM_ANDINFO; |
| 886 | pOrTerm->eOperator = WO_AND; |
| 887 | pAndWC = &pAndInfo->wc; |
| 888 | whereClauseInit(pAndWC, pWC->pParse, pMaskSet); |
| 889 | whereSplit(pAndWC, pOrTerm->pExpr, TK_AND); |
| 890 | exprAnalyzeAll(pSrc, pAndWC); |
drh | 7c2fbde | 2009-01-07 20:58:57 +0000 | [diff] [blame] | 891 | testcase( db->mallocFailed ); |
drh | 96c7a7d | 2009-01-10 15:34:12 +0000 | [diff] [blame] | 892 | if( !db->mallocFailed ){ |
| 893 | for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){ |
| 894 | assert( pAndTerm->pExpr ); |
| 895 | if( allowedOp(pAndTerm->pExpr->op) ){ |
| 896 | b |= getMask(pMaskSet, pAndTerm->leftCursor); |
| 897 | } |
drh | 2943525 | 2008-12-28 18:35:08 +0000 | [diff] [blame] | 898 | } |
| 899 | } |
| 900 | indexable &= b; |
| 901 | } |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 902 | }else if( pOrTerm->wtFlags & TERM_COPIED ){ |
| 903 | /* Skip this term for now. We revisit it when we process the |
| 904 | ** corresponding TERM_VIRTUAL term */ |
| 905 | }else{ |
| 906 | Bitmask b; |
| 907 | b = getMask(pMaskSet, pOrTerm->leftCursor); |
| 908 | if( pOrTerm->wtFlags & TERM_VIRTUAL ){ |
| 909 | WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent]; |
| 910 | b |= getMask(pMaskSet, pOther->leftCursor); |
| 911 | } |
| 912 | indexable &= b; |
| 913 | if( pOrTerm->eOperator!=WO_EQ ){ |
| 914 | chngToIN = 0; |
| 915 | }else{ |
| 916 | chngToIN &= b; |
| 917 | } |
| 918 | } |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 919 | } |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 920 | |
| 921 | /* |
| 922 | ** Record the set of tables that satisfy case 2. The set might be |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 923 | ** empty. |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 924 | */ |
| 925 | pOrInfo->indexable = indexable; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 926 | pTerm->eOperator = indexable==0 ? 0 : WO_OR; |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 927 | |
| 928 | /* |
| 929 | ** chngToIN holds a set of tables that *might* satisfy case 1. But |
| 930 | ** we have to do some additional checking to see if case 1 really |
| 931 | ** is satisfied. |
drh | 4e8be3b | 2009-06-08 17:11:08 +0000 | [diff] [blame] | 932 | ** |
| 933 | ** chngToIN will hold either 0, 1, or 2 bits. The 0-bit case means |
| 934 | ** that there is no possibility of transforming the OR clause into an |
| 935 | ** IN operator because one or more terms in the OR clause contain |
| 936 | ** something other than == on a column in the single table. The 1-bit |
| 937 | ** case means that every term of the OR clause is of the form |
| 938 | ** "table.column=expr" for some single table. The one bit that is set |
| 939 | ** will correspond to the common table. We still need to check to make |
| 940 | ** sure the same column is used on all terms. The 2-bit case is when |
| 941 | ** the all terms are of the form "table1.column=table2.column". It |
| 942 | ** might be possible to form an IN operator with either table1.column |
| 943 | ** or table2.column as the LHS if either is common to every term of |
| 944 | ** the OR clause. |
| 945 | ** |
| 946 | ** Note that terms of the form "table.column1=table.column2" (the |
| 947 | ** same table on both sizes of the ==) cannot be optimized. |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 948 | */ |
| 949 | if( chngToIN ){ |
| 950 | int okToChngToIN = 0; /* True if the conversion to IN is valid */ |
| 951 | int iColumn = -1; /* Column index on lhs of IN operator */ |
shane | 63207ab | 2009-02-04 01:49:30 +0000 | [diff] [blame] | 952 | int iCursor = -1; /* Table cursor common to all terms */ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 953 | int j = 0; /* Loop counter */ |
| 954 | |
| 955 | /* Search for a table and column that appears on one side or the |
| 956 | ** other of the == operator in every subterm. That table and column |
| 957 | ** will be recorded in iCursor and iColumn. There might not be any |
| 958 | ** such table and column. Set okToChngToIN if an appropriate table |
| 959 | ** and column is found but leave okToChngToIN false if not found. |
| 960 | */ |
| 961 | for(j=0; j<2 && !okToChngToIN; j++){ |
| 962 | pOrTerm = pOrWc->a; |
| 963 | for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){ |
| 964 | assert( pOrTerm->eOperator==WO_EQ ); |
| 965 | pOrTerm->wtFlags &= ~TERM_OR_OK; |
drh | 4e8be3b | 2009-06-08 17:11:08 +0000 | [diff] [blame] | 966 | if( pOrTerm->leftCursor==iCursor ){ |
| 967 | /* This is the 2-bit case and we are on the second iteration and |
| 968 | ** current term is from the first iteration. So skip this term. */ |
| 969 | assert( j==1 ); |
| 970 | continue; |
| 971 | } |
| 972 | if( (chngToIN & getMask(pMaskSet, pOrTerm->leftCursor))==0 ){ |
| 973 | /* This term must be of the form t1.a==t2.b where t2 is in the |
| 974 | ** chngToIN set but t1 is not. This term will be either preceeded |
| 975 | ** or follwed by an inverted copy (t2.b==t1.a). Skip this term |
| 976 | ** and use its inversion. */ |
| 977 | testcase( pOrTerm->wtFlags & TERM_COPIED ); |
| 978 | testcase( pOrTerm->wtFlags & TERM_VIRTUAL ); |
| 979 | assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) ); |
| 980 | continue; |
| 981 | } |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 982 | iColumn = pOrTerm->u.leftColumn; |
| 983 | iCursor = pOrTerm->leftCursor; |
| 984 | break; |
| 985 | } |
| 986 | if( i<0 ){ |
drh | 4e8be3b | 2009-06-08 17:11:08 +0000 | [diff] [blame] | 987 | /* No candidate table+column was found. This can only occur |
| 988 | ** on the second iteration */ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 989 | assert( j==1 ); |
| 990 | assert( (chngToIN&(chngToIN-1))==0 ); |
drh | 4e8be3b | 2009-06-08 17:11:08 +0000 | [diff] [blame] | 991 | assert( chngToIN==getMask(pMaskSet, iCursor) ); |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 992 | break; |
| 993 | } |
drh | 4e8be3b | 2009-06-08 17:11:08 +0000 | [diff] [blame] | 994 | testcase( j==1 ); |
| 995 | |
| 996 | /* We have found a candidate table and column. Check to see if that |
| 997 | ** table and column is common to every term in the OR clause */ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 998 | okToChngToIN = 1; |
| 999 | for(; i>=0 && okToChngToIN; i--, pOrTerm++){ |
| 1000 | assert( pOrTerm->eOperator==WO_EQ ); |
| 1001 | if( pOrTerm->leftCursor!=iCursor ){ |
| 1002 | pOrTerm->wtFlags &= ~TERM_OR_OK; |
| 1003 | }else if( pOrTerm->u.leftColumn!=iColumn ){ |
| 1004 | okToChngToIN = 0; |
| 1005 | }else{ |
| 1006 | int affLeft, affRight; |
| 1007 | /* If the right-hand side is also a column, then the affinities |
| 1008 | ** of both right and left sides must be such that no type |
| 1009 | ** conversions are required on the right. (Ticket #2249) |
| 1010 | */ |
| 1011 | affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight); |
| 1012 | affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft); |
| 1013 | if( affRight!=0 && affRight!=affLeft ){ |
| 1014 | okToChngToIN = 0; |
| 1015 | }else{ |
| 1016 | pOrTerm->wtFlags |= TERM_OR_OK; |
| 1017 | } |
| 1018 | } |
| 1019 | } |
| 1020 | } |
| 1021 | |
| 1022 | /* At this point, okToChngToIN is true if original pTerm satisfies |
| 1023 | ** case 1. In that case, construct a new virtual term that is |
| 1024 | ** pTerm converted into an IN operator. |
| 1025 | */ |
| 1026 | if( okToChngToIN ){ |
| 1027 | Expr *pDup; /* A transient duplicate expression */ |
| 1028 | ExprList *pList = 0; /* The RHS of the IN operator */ |
| 1029 | Expr *pLeft = 0; /* The LHS of the IN operator */ |
| 1030 | Expr *pNew; /* The complete IN operator */ |
| 1031 | |
| 1032 | for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){ |
| 1033 | if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue; |
| 1034 | assert( pOrTerm->eOperator==WO_EQ ); |
| 1035 | assert( pOrTerm->leftCursor==iCursor ); |
| 1036 | assert( pOrTerm->u.leftColumn==iColumn ); |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1037 | pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0); |
drh | b7916a7 | 2009-05-27 10:31:29 +0000 | [diff] [blame] | 1038 | pList = sqlite3ExprListAppend(pWC->pParse, pList, pDup); |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1039 | pLeft = pOrTerm->pExpr->pLeft; |
| 1040 | } |
| 1041 | assert( pLeft!=0 ); |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1042 | pDup = sqlite3ExprDup(db, pLeft, 0); |
drh | b7916a7 | 2009-05-27 10:31:29 +0000 | [diff] [blame] | 1043 | pNew = sqlite3PExpr(pParse, TK_IN, pDup, 0, 0); |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1044 | if( pNew ){ |
| 1045 | int idxNew; |
| 1046 | transferJoinMarkings(pNew, pExpr); |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1047 | assert( !ExprHasProperty(pNew, EP_xIsSelect) ); |
| 1048 | pNew->x.pList = pList; |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1049 | idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC); |
| 1050 | testcase( idxNew==0 ); |
| 1051 | exprAnalyze(pSrc, pWC, idxNew); |
| 1052 | pTerm = &pWC->a[idxTerm]; |
| 1053 | pWC->a[idxNew].iParent = idxTerm; |
| 1054 | pTerm->nChild = 1; |
| 1055 | }else{ |
| 1056 | sqlite3ExprListDelete(db, pList); |
| 1057 | } |
| 1058 | pTerm->eOperator = 0; /* case 1 trumps case 2 */ |
| 1059 | } |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 1060 | } |
drh | 3e35580 | 2007-02-23 23:13:33 +0000 | [diff] [blame] | 1061 | } |
| 1062 | #endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */ |
drh | 54a167d | 2005-11-26 14:08:07 +0000 | [diff] [blame] | 1063 | |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1064 | |
drh | 54a167d | 2005-11-26 14:08:07 +0000 | [diff] [blame] | 1065 | /* |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 1066 | ** The input to this routine is an WhereTerm structure with only the |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 1067 | ** "pExpr" field filled in. The job of this routine is to analyze the |
drh | 0aa74ed | 2005-07-16 13:33:20 +0000 | [diff] [blame] | 1068 | ** subexpression and populate all the other fields of the WhereTerm |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 1069 | ** structure. |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 1070 | ** |
| 1071 | ** If the expression is of the form "<expr> <op> X" it gets commuted |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1072 | ** to the standard form of "X <op> <expr>". |
| 1073 | ** |
| 1074 | ** If the expression is of the form "X <op> Y" where both X and Y are |
| 1075 | ** columns, then the original expression is unchanged and a new virtual |
| 1076 | ** term of the form "Y <op> X" is added to the WHERE clause and |
| 1077 | ** analyzed separately. The original term is marked with TERM_COPIED |
| 1078 | ** and the new term is marked with TERM_DYNAMIC (because it's pExpr |
| 1079 | ** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it |
| 1080 | ** is a commuted copy of a prior term.) The original term has nChild=1 |
| 1081 | ** and the copy has idxParent set to the index of the original term. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 1082 | */ |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1083 | static void exprAnalyze( |
| 1084 | SrcList *pSrc, /* the FROM clause */ |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1085 | WhereClause *pWC, /* the WHERE clause */ |
| 1086 | int idxTerm /* Index of the term to be analyzed */ |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1087 | ){ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1088 | WhereTerm *pTerm; /* The term to be analyzed */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 1089 | WhereMaskSet *pMaskSet; /* Set of table index masks */ |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1090 | Expr *pExpr; /* The expression to be analyzed */ |
| 1091 | Bitmask prereqLeft; /* Prerequesites of the pExpr->pLeft */ |
| 1092 | Bitmask prereqAll; /* Prerequesites of pExpr */ |
drh | dafc0ce | 2008-04-17 19:14:02 +0000 | [diff] [blame] | 1093 | Bitmask extraRight = 0; |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1094 | int isComplete; |
drh | 9f504ea | 2008-02-23 21:55:39 +0000 | [diff] [blame] | 1095 | int noCase; |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1096 | int op; /* Top-level operator. pExpr->op */ |
| 1097 | Parse *pParse = pWC->pParse; /* Parsing context */ |
| 1098 | sqlite3 *db = pParse->db; /* Database connection */ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 1099 | Expr *pStr1; |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1100 | |
drh | f998b73 | 2007-11-26 13:36:00 +0000 | [diff] [blame] | 1101 | if( db->mallocFailed ){ |
| 1102 | return; |
| 1103 | } |
| 1104 | pTerm = &pWC->a[idxTerm]; |
| 1105 | pMaskSet = pWC->pMaskSet; |
| 1106 | pExpr = pTerm->pExpr; |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1107 | prereqLeft = exprTableUsage(pMaskSet, pExpr->pLeft); |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 1108 | op = pExpr->op; |
| 1109 | if( op==TK_IN ){ |
drh | f5b1138 | 2005-09-17 13:07:13 +0000 | [diff] [blame] | 1110 | assert( pExpr->pRight==0 ); |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1111 | if( ExprHasProperty(pExpr, EP_xIsSelect) ){ |
| 1112 | pTerm->prereqRight = exprSelectTableUsage(pMaskSet, pExpr->x.pSelect); |
| 1113 | }else{ |
| 1114 | pTerm->prereqRight = exprListTableUsage(pMaskSet, pExpr->x.pList); |
| 1115 | } |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 1116 | }else if( op==TK_ISNULL ){ |
| 1117 | pTerm->prereqRight = 0; |
drh | f5b1138 | 2005-09-17 13:07:13 +0000 | [diff] [blame] | 1118 | }else{ |
| 1119 | pTerm->prereqRight = exprTableUsage(pMaskSet, pExpr->pRight); |
| 1120 | } |
drh | 22d6a53 | 2005-09-19 21:05:48 +0000 | [diff] [blame] | 1121 | prereqAll = exprTableUsage(pMaskSet, pExpr); |
| 1122 | if( ExprHasProperty(pExpr, EP_FromJoin) ){ |
drh | 42165be | 2008-03-26 14:56:34 +0000 | [diff] [blame] | 1123 | Bitmask x = getMask(pMaskSet, pExpr->iRightJoinTable); |
| 1124 | prereqAll |= x; |
drh | dafc0ce | 2008-04-17 19:14:02 +0000 | [diff] [blame] | 1125 | extraRight = x-1; /* ON clause terms may not be used with an index |
| 1126 | ** on left table of a LEFT JOIN. Ticket #3015 */ |
drh | 22d6a53 | 2005-09-19 21:05:48 +0000 | [diff] [blame] | 1127 | } |
| 1128 | pTerm->prereqAll = prereqAll; |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1129 | pTerm->leftCursor = -1; |
drh | 45b1ee4 | 2005-08-02 17:48:22 +0000 | [diff] [blame] | 1130 | pTerm->iParent = -1; |
drh | b52076c | 2006-01-23 13:22:09 +0000 | [diff] [blame] | 1131 | pTerm->eOperator = 0; |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 1132 | if( allowedOp(op) && (pTerm->prereqRight & prereqLeft)==0 ){ |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1133 | Expr *pLeft = pExpr->pLeft; |
| 1134 | Expr *pRight = pExpr->pRight; |
| 1135 | if( pLeft->op==TK_COLUMN ){ |
| 1136 | pTerm->leftCursor = pLeft->iTable; |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 1137 | pTerm->u.leftColumn = pLeft->iColumn; |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 1138 | pTerm->eOperator = operatorMask(op); |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 1139 | } |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1140 | if( pRight && pRight->op==TK_COLUMN ){ |
| 1141 | WhereTerm *pNew; |
| 1142 | Expr *pDup; |
| 1143 | if( pTerm->leftCursor>=0 ){ |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1144 | int idxNew; |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1145 | pDup = sqlite3ExprDup(db, pExpr, 0); |
drh | 1743575 | 2007-08-16 04:30:38 +0000 | [diff] [blame] | 1146 | if( db->mallocFailed ){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 1147 | sqlite3ExprDelete(db, pDup); |
drh | 28f4591 | 2006-10-18 23:26:38 +0000 | [diff] [blame] | 1148 | return; |
| 1149 | } |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1150 | idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC); |
| 1151 | if( idxNew==0 ) return; |
| 1152 | pNew = &pWC->a[idxNew]; |
| 1153 | pNew->iParent = idxTerm; |
| 1154 | pTerm = &pWC->a[idxTerm]; |
drh | 45b1ee4 | 2005-08-02 17:48:22 +0000 | [diff] [blame] | 1155 | pTerm->nChild = 1; |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 1156 | pTerm->wtFlags |= TERM_COPIED; |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1157 | }else{ |
| 1158 | pDup = pExpr; |
| 1159 | pNew = pTerm; |
| 1160 | } |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 1161 | exprCommute(pParse, pDup); |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1162 | pLeft = pDup->pLeft; |
| 1163 | pNew->leftCursor = pLeft->iTable; |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 1164 | pNew->u.leftColumn = pLeft->iColumn; |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1165 | pNew->prereqRight = prereqLeft; |
| 1166 | pNew->prereqAll = prereqAll; |
drh | b52076c | 2006-01-23 13:22:09 +0000 | [diff] [blame] | 1167 | pNew->eOperator = operatorMask(pDup->op); |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 1168 | } |
| 1169 | } |
drh | ed37800 | 2005-07-28 23:12:08 +0000 | [diff] [blame] | 1170 | |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1171 | #ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION |
drh | ed37800 | 2005-07-28 23:12:08 +0000 | [diff] [blame] | 1172 | /* If a term is the BETWEEN operator, create two new virtual terms |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1173 | ** that define the range that the BETWEEN implements. For example: |
| 1174 | ** |
| 1175 | ** a BETWEEN b AND c |
| 1176 | ** |
| 1177 | ** is converted into: |
| 1178 | ** |
| 1179 | ** (a BETWEEN b AND c) AND (a>=b) AND (a<=c) |
| 1180 | ** |
| 1181 | ** The two new terms are added onto the end of the WhereClause object. |
| 1182 | ** The new terms are "dynamic" and are children of the original BETWEEN |
| 1183 | ** term. That means that if the BETWEEN term is coded, the children are |
| 1184 | ** skipped. Or, if the children are satisfied by an index, the original |
| 1185 | ** BETWEEN term is skipped. |
drh | ed37800 | 2005-07-28 23:12:08 +0000 | [diff] [blame] | 1186 | */ |
drh | 2943525 | 2008-12-28 18:35:08 +0000 | [diff] [blame] | 1187 | else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){ |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1188 | ExprList *pList = pExpr->x.pList; |
drh | ed37800 | 2005-07-28 23:12:08 +0000 | [diff] [blame] | 1189 | int i; |
| 1190 | static const u8 ops[] = {TK_GE, TK_LE}; |
| 1191 | assert( pList!=0 ); |
| 1192 | assert( pList->nExpr==2 ); |
| 1193 | for(i=0; i<2; i++){ |
| 1194 | Expr *pNewExpr; |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1195 | int idxNew; |
drh | b7916a7 | 2009-05-27 10:31:29 +0000 | [diff] [blame] | 1196 | pNewExpr = sqlite3PExpr(pParse, ops[i], |
| 1197 | sqlite3ExprDup(db, pExpr->pLeft, 0), |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1198 | sqlite3ExprDup(db, pList->a[i].pExpr, 0), 0); |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1199 | idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC); |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 1200 | testcase( idxNew==0 ); |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1201 | exprAnalyze(pSrc, pWC, idxNew); |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1202 | pTerm = &pWC->a[idxTerm]; |
| 1203 | pWC->a[idxNew].iParent = idxTerm; |
drh | ed37800 | 2005-07-28 23:12:08 +0000 | [diff] [blame] | 1204 | } |
drh | 45b1ee4 | 2005-08-02 17:48:22 +0000 | [diff] [blame] | 1205 | pTerm->nChild = 2; |
drh | ed37800 | 2005-07-28 23:12:08 +0000 | [diff] [blame] | 1206 | } |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1207 | #endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */ |
drh | ed37800 | 2005-07-28 23:12:08 +0000 | [diff] [blame] | 1208 | |
danielk1977 | 1576cd9 | 2006-01-14 08:02:28 +0000 | [diff] [blame] | 1209 | #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY) |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1210 | /* Analyze a term that is composed of two or more subterms connected by |
| 1211 | ** an OR operator. |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 1212 | */ |
| 1213 | else if( pExpr->op==TK_OR ){ |
drh | 2943525 | 2008-12-28 18:35:08 +0000 | [diff] [blame] | 1214 | assert( pWC->op==TK_AND ); |
drh | 1a58fe0 | 2008-12-20 02:06:13 +0000 | [diff] [blame] | 1215 | exprAnalyzeOrTerm(pSrc, pWC, idxTerm); |
danielk1977 | f51d1bd | 2009-07-31 06:14:51 +0000 | [diff] [blame] | 1216 | pTerm = &pWC->a[idxTerm]; |
drh | 6c30be8 | 2005-07-29 15:10:17 +0000 | [diff] [blame] | 1217 | } |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1218 | #endif /* SQLITE_OMIT_OR_OPTIMIZATION */ |
| 1219 | |
| 1220 | #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION |
| 1221 | /* Add constraints to reduce the search space on a LIKE or GLOB |
| 1222 | ** operator. |
drh | 9f504ea | 2008-02-23 21:55:39 +0000 | [diff] [blame] | 1223 | ** |
| 1224 | ** A like pattern of the form "x LIKE 'abc%'" is changed into constraints |
| 1225 | ** |
| 1226 | ** x>='abc' AND x<'abd' AND x LIKE 'abc%' |
| 1227 | ** |
| 1228 | ** The last character of the prefix "abc" is incremented to form the |
shane | 7bc71e5 | 2008-05-28 18:01:44 +0000 | [diff] [blame] | 1229 | ** termination condition "abd". |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1230 | */ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 1231 | if( pWC->op==TK_AND |
| 1232 | && isLikeOrGlob(pParse, pExpr, &pStr1, &isComplete, &noCase) |
| 1233 | ){ |
| 1234 | Expr *pLeft; |
| 1235 | Expr *pStr2; |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1236 | Expr *pNewExpr1, *pNewExpr2; |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1237 | int idxNew1, idxNew2; |
| 1238 | |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1239 | pLeft = pExpr->x.pList->a[1].pExpr; |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1240 | pStr2 = sqlite3ExprDup(db, pStr1, 0); |
drh | f998b73 | 2007-11-26 13:36:00 +0000 | [diff] [blame] | 1241 | if( !db->mallocFailed ){ |
drh | 254993e | 2009-06-08 19:44:36 +0000 | [diff] [blame] | 1242 | u8 c, *pC; /* Last character before the first wildcard */ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 1243 | pC = (u8*)&pStr2->u.zToken[sqlite3Strlen30(pStr2->u.zToken)-1]; |
drh | 9f504ea | 2008-02-23 21:55:39 +0000 | [diff] [blame] | 1244 | c = *pC; |
drh | 02a50b7 | 2008-05-26 18:33:40 +0000 | [diff] [blame] | 1245 | if( noCase ){ |
drh | 254993e | 2009-06-08 19:44:36 +0000 | [diff] [blame] | 1246 | /* The point is to increment the last character before the first |
| 1247 | ** wildcard. But if we increment '@', that will push it into the |
| 1248 | ** alphabetic range where case conversions will mess up the |
| 1249 | ** inequality. To avoid this, make sure to also run the full |
| 1250 | ** LIKE on all candidate expressions by clearing the isComplete flag |
| 1251 | */ |
| 1252 | if( c=='A'-1 ) isComplete = 0; |
| 1253 | |
drh | 02a50b7 | 2008-05-26 18:33:40 +0000 | [diff] [blame] | 1254 | c = sqlite3UpperToLower[c]; |
| 1255 | } |
drh | 9f504ea | 2008-02-23 21:55:39 +0000 | [diff] [blame] | 1256 | *pC = c + 1; |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1257 | } |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1258 | pNewExpr1 = sqlite3PExpr(pParse, TK_GE, sqlite3ExprDup(db,pLeft,0),pStr1,0); |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1259 | idxNew1 = whereClauseInsert(pWC, pNewExpr1, TERM_VIRTUAL|TERM_DYNAMIC); |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 1260 | testcase( idxNew1==0 ); |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1261 | exprAnalyze(pSrc, pWC, idxNew1); |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1262 | pNewExpr2 = sqlite3PExpr(pParse, TK_LT, sqlite3ExprDup(db,pLeft,0),pStr2,0); |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1263 | idxNew2 = whereClauseInsert(pWC, pNewExpr2, TERM_VIRTUAL|TERM_DYNAMIC); |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 1264 | testcase( idxNew2==0 ); |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1265 | exprAnalyze(pSrc, pWC, idxNew2); |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1266 | pTerm = &pWC->a[idxTerm]; |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1267 | if( isComplete ){ |
drh | 9eb2028 | 2005-08-24 03:52:18 +0000 | [diff] [blame] | 1268 | pWC->a[idxNew1].iParent = idxTerm; |
| 1269 | pWC->a[idxNew2].iParent = idxTerm; |
drh | d2687b7 | 2005-08-12 22:56:09 +0000 | [diff] [blame] | 1270 | pTerm->nChild = 2; |
| 1271 | } |
| 1272 | } |
| 1273 | #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */ |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1274 | |
| 1275 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
| 1276 | /* Add a WO_MATCH auxiliary term to the constraint set if the |
| 1277 | ** current expression is of the form: column MATCH expr. |
| 1278 | ** This information is used by the xBestIndex methods of |
| 1279 | ** virtual tables. The native query optimizer does not attempt |
| 1280 | ** to do anything with MATCH functions. |
| 1281 | */ |
| 1282 | if( isMatchOfColumn(pExpr) ){ |
| 1283 | int idxNew; |
| 1284 | Expr *pRight, *pLeft; |
| 1285 | WhereTerm *pNewTerm; |
| 1286 | Bitmask prereqColumn, prereqExpr; |
| 1287 | |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 1288 | pRight = pExpr->x.pList->a[0].pExpr; |
| 1289 | pLeft = pExpr->x.pList->a[1].pExpr; |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1290 | prereqExpr = exprTableUsage(pMaskSet, pRight); |
| 1291 | prereqColumn = exprTableUsage(pMaskSet, pLeft); |
| 1292 | if( (prereqExpr & prereqColumn)==0 ){ |
drh | 1a90e09 | 2006-06-14 22:07:10 +0000 | [diff] [blame] | 1293 | Expr *pNewExpr; |
drh | b7916a7 | 2009-05-27 10:31:29 +0000 | [diff] [blame] | 1294 | pNewExpr = sqlite3PExpr(pParse, TK_MATCH, |
| 1295 | 0, sqlite3ExprDup(db, pRight, 0), 0); |
drh | 1a90e09 | 2006-06-14 22:07:10 +0000 | [diff] [blame] | 1296 | idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC); |
drh | 6a1e071 | 2008-12-05 15:24:15 +0000 | [diff] [blame] | 1297 | testcase( idxNew==0 ); |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1298 | pNewTerm = &pWC->a[idxNew]; |
| 1299 | pNewTerm->prereqRight = prereqExpr; |
| 1300 | pNewTerm->leftCursor = pLeft->iTable; |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 1301 | pNewTerm->u.leftColumn = pLeft->iColumn; |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1302 | pNewTerm->eOperator = WO_MATCH; |
| 1303 | pNewTerm->iParent = idxTerm; |
drh | d2ca60d | 2006-06-27 02:36:58 +0000 | [diff] [blame] | 1304 | pTerm = &pWC->a[idxTerm]; |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1305 | pTerm->nChild = 1; |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 1306 | pTerm->wtFlags |= TERM_COPIED; |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1307 | pNewTerm->prereqAll = pTerm->prereqAll; |
| 1308 | } |
| 1309 | } |
| 1310 | #endif /* SQLITE_OMIT_VIRTUALTABLE */ |
drh | dafc0ce | 2008-04-17 19:14:02 +0000 | [diff] [blame] | 1311 | |
| 1312 | /* Prevent ON clause terms of a LEFT JOIN from being used to drive |
| 1313 | ** an index for tables to the left of the join. |
| 1314 | */ |
| 1315 | pTerm->prereqRight |= extraRight; |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 1316 | } |
| 1317 | |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1318 | /* |
| 1319 | ** Return TRUE if any of the expressions in pList->a[iFirst...] contain |
| 1320 | ** a reference to any table other than the iBase table. |
| 1321 | */ |
| 1322 | static int referencesOtherTables( |
| 1323 | ExprList *pList, /* Search expressions in ths list */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 1324 | WhereMaskSet *pMaskSet, /* Mapping from tables to bitmaps */ |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1325 | int iFirst, /* Be searching with the iFirst-th expression */ |
| 1326 | int iBase /* Ignore references to this table */ |
| 1327 | ){ |
| 1328 | Bitmask allowed = ~getMask(pMaskSet, iBase); |
| 1329 | while( iFirst<pList->nExpr ){ |
| 1330 | if( (exprTableUsage(pMaskSet, pList->a[iFirst++].pExpr)&allowed)!=0 ){ |
| 1331 | return 1; |
| 1332 | } |
| 1333 | } |
| 1334 | return 0; |
| 1335 | } |
| 1336 | |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 1337 | |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 1338 | /* |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1339 | ** This routine decides if pIdx can be used to satisfy the ORDER BY |
| 1340 | ** clause. If it can, it returns 1. If pIdx cannot satisfy the |
| 1341 | ** ORDER BY clause, this routine returns 0. |
| 1342 | ** |
| 1343 | ** pOrderBy is an ORDER BY clause from a SELECT statement. pTab is the |
| 1344 | ** left-most table in the FROM clause of that same SELECT statement and |
| 1345 | ** the table has a cursor number of "base". pIdx is an index on pTab. |
| 1346 | ** |
| 1347 | ** nEqCol is the number of columns of pIdx that are used as equality |
| 1348 | ** constraints. Any of these columns may be missing from the ORDER BY |
| 1349 | ** clause and the match can still be a success. |
| 1350 | ** |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1351 | ** All terms of the ORDER BY that match against the index must be either |
| 1352 | ** ASC or DESC. (Terms of the ORDER BY clause past the end of a UNIQUE |
| 1353 | ** index do not need to satisfy this constraint.) The *pbRev value is |
| 1354 | ** set to 1 if the ORDER BY clause is all DESC and it is set to 0 if |
| 1355 | ** the ORDER BY clause is all ASC. |
| 1356 | */ |
| 1357 | static int isSortingIndex( |
| 1358 | Parse *pParse, /* Parsing context */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 1359 | WhereMaskSet *pMaskSet, /* Mapping from table cursor numbers to bitmaps */ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1360 | Index *pIdx, /* The index we are testing */ |
drh | 7416170 | 2006-02-24 02:53:49 +0000 | [diff] [blame] | 1361 | int base, /* Cursor number for the table to be sorted */ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1362 | ExprList *pOrderBy, /* The ORDER BY clause */ |
| 1363 | int nEqCol, /* Number of index columns with == constraints */ |
| 1364 | int *pbRev /* Set to 1 if ORDER BY is DESC */ |
| 1365 | ){ |
drh | b46b577 | 2005-08-29 16:40:52 +0000 | [diff] [blame] | 1366 | int i, j; /* Loop counters */ |
drh | 85eeb69 | 2005-12-21 03:16:42 +0000 | [diff] [blame] | 1367 | int sortOrder = 0; /* XOR of index and ORDER BY sort direction */ |
drh | b46b577 | 2005-08-29 16:40:52 +0000 | [diff] [blame] | 1368 | int nTerm; /* Number of ORDER BY terms */ |
| 1369 | struct ExprList_item *pTerm; /* A term of the ORDER BY clause */ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1370 | sqlite3 *db = pParse->db; |
| 1371 | |
| 1372 | assert( pOrderBy!=0 ); |
| 1373 | nTerm = pOrderBy->nExpr; |
| 1374 | assert( nTerm>0 ); |
| 1375 | |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 1376 | /* Argument pIdx must either point to a 'real' named index structure, |
| 1377 | ** or an index structure allocated on the stack by bestBtreeIndex() to |
| 1378 | ** represent the rowid index that is part of every table. */ |
| 1379 | assert( pIdx->zName || (pIdx->nColumn==1 && pIdx->aiColumn[0]==-1) ); |
| 1380 | |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1381 | /* Match terms of the ORDER BY clause against columns of |
| 1382 | ** the index. |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1383 | ** |
| 1384 | ** Note that indices have pIdx->nColumn regular columns plus |
| 1385 | ** one additional column containing the rowid. The rowid column |
| 1386 | ** of the index is also allowed to match against the ORDER BY |
| 1387 | ** clause. |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1388 | */ |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1389 | for(i=j=0, pTerm=pOrderBy->a; j<nTerm && i<=pIdx->nColumn; i++){ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1390 | Expr *pExpr; /* The expression of the ORDER BY pTerm */ |
| 1391 | CollSeq *pColl; /* The collating sequence of pExpr */ |
drh | 85eeb69 | 2005-12-21 03:16:42 +0000 | [diff] [blame] | 1392 | int termSortOrder; /* Sort order for this term */ |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1393 | int iColumn; /* The i-th column of the index. -1 for rowid */ |
| 1394 | int iSortOrder; /* 1 for DESC, 0 for ASC on the i-th index term */ |
| 1395 | const char *zColl; /* Name of the collating sequence for i-th index term */ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1396 | |
| 1397 | pExpr = pTerm->pExpr; |
| 1398 | if( pExpr->op!=TK_COLUMN || pExpr->iTable!=base ){ |
| 1399 | /* Can not use an index sort on anything that is not a column in the |
| 1400 | ** left-most table of the FROM clause */ |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1401 | break; |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1402 | } |
| 1403 | pColl = sqlite3ExprCollSeq(pParse, pExpr); |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1404 | if( !pColl ){ |
| 1405 | pColl = db->pDfltColl; |
| 1406 | } |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 1407 | if( pIdx->zName && i<pIdx->nColumn ){ |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1408 | iColumn = pIdx->aiColumn[i]; |
| 1409 | if( iColumn==pIdx->pTable->iPKey ){ |
| 1410 | iColumn = -1; |
| 1411 | } |
| 1412 | iSortOrder = pIdx->aSortOrder[i]; |
| 1413 | zColl = pIdx->azColl[i]; |
| 1414 | }else{ |
| 1415 | iColumn = -1; |
| 1416 | iSortOrder = 0; |
| 1417 | zColl = pColl->zName; |
| 1418 | } |
| 1419 | if( pExpr->iColumn!=iColumn || sqlite3StrICmp(pColl->zName, zColl) ){ |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 1420 | /* Term j of the ORDER BY clause does not match column i of the index */ |
| 1421 | if( i<nEqCol ){ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1422 | /* If an index column that is constrained by == fails to match an |
| 1423 | ** ORDER BY term, that is OK. Just ignore that column of the index |
| 1424 | */ |
| 1425 | continue; |
drh | ff354e9 | 2008-06-25 02:47:57 +0000 | [diff] [blame] | 1426 | }else if( i==pIdx->nColumn ){ |
| 1427 | /* Index column i is the rowid. All other terms match. */ |
| 1428 | break; |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1429 | }else{ |
| 1430 | /* If an index column fails to match and is not constrained by == |
| 1431 | ** then the index cannot satisfy the ORDER BY constraint. |
| 1432 | */ |
| 1433 | return 0; |
| 1434 | } |
| 1435 | } |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 1436 | assert( pIdx->aSortOrder!=0 || iColumn==-1 ); |
drh | 85eeb69 | 2005-12-21 03:16:42 +0000 | [diff] [blame] | 1437 | assert( pTerm->sortOrder==0 || pTerm->sortOrder==1 ); |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1438 | assert( iSortOrder==0 || iSortOrder==1 ); |
| 1439 | termSortOrder = iSortOrder ^ pTerm->sortOrder; |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1440 | if( i>nEqCol ){ |
drh | 85eeb69 | 2005-12-21 03:16:42 +0000 | [diff] [blame] | 1441 | if( termSortOrder!=sortOrder ){ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1442 | /* Indices can only be used if all ORDER BY terms past the |
| 1443 | ** equality constraints are all either DESC or ASC. */ |
| 1444 | return 0; |
| 1445 | } |
| 1446 | }else{ |
drh | 85eeb69 | 2005-12-21 03:16:42 +0000 | [diff] [blame] | 1447 | sortOrder = termSortOrder; |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1448 | } |
| 1449 | j++; |
| 1450 | pTerm++; |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1451 | if( iColumn<0 && !referencesOtherTables(pOrderBy, pMaskSet, j, base) ){ |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1452 | /* If the indexed column is the primary key and everything matches |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1453 | ** so far and none of the ORDER BY terms to the right reference other |
| 1454 | ** tables in the join, then we are assured that the index can be used |
| 1455 | ** to sort because the primary key is unique and so none of the other |
| 1456 | ** columns will make any difference |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1457 | */ |
| 1458 | j = nTerm; |
| 1459 | } |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1460 | } |
| 1461 | |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1462 | *pbRev = sortOrder!=0; |
drh | 8718f52 | 2005-08-13 16:13:04 +0000 | [diff] [blame] | 1463 | if( j>=nTerm ){ |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1464 | /* All terms of the ORDER BY clause are covered by this index so |
| 1465 | ** this index can be used for sorting. */ |
| 1466 | return 1; |
| 1467 | } |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1468 | if( pIdx->onError!=OE_None && i==pIdx->nColumn |
| 1469 | && !referencesOtherTables(pOrderBy, pMaskSet, j, base) ){ |
drh | cc19254 | 2006-12-20 03:24:19 +0000 | [diff] [blame] | 1470 | /* All terms of this index match some prefix of the ORDER BY clause |
drh | 7b4fc6a | 2007-02-06 13:26:32 +0000 | [diff] [blame] | 1471 | ** and the index is UNIQUE and no terms on the tail of the ORDER BY |
| 1472 | ** clause reference other tables in a join. If this is all true then |
| 1473 | ** the order by clause is superfluous. */ |
drh | 5166986 | 2004-12-18 18:40:26 +0000 | [diff] [blame] | 1474 | return 1; |
| 1475 | } |
| 1476 | return 0; |
| 1477 | } |
| 1478 | |
| 1479 | /* |
drh | b6fb62d | 2005-09-20 08:47:20 +0000 | [diff] [blame] | 1480 | ** Prepare a crude estimate of the logarithm of the input value. |
drh | 28c4cf4 | 2005-07-27 20:41:43 +0000 | [diff] [blame] | 1481 | ** The results need not be exact. This is only used for estimating |
drh | 909626d | 2008-05-30 14:58:37 +0000 | [diff] [blame] | 1482 | ** the total cost of performing operations with O(logN) or O(NlogN) |
drh | 28c4cf4 | 2005-07-27 20:41:43 +0000 | [diff] [blame] | 1483 | ** complexity. Because N is just a guess, it is no great tragedy if |
| 1484 | ** logN is a little off. |
drh | 28c4cf4 | 2005-07-27 20:41:43 +0000 | [diff] [blame] | 1485 | */ |
| 1486 | static double estLog(double N){ |
drh | b37df7b | 2005-10-13 02:09:49 +0000 | [diff] [blame] | 1487 | double logN = 1; |
| 1488 | double x = 10; |
drh | 28c4cf4 | 2005-07-27 20:41:43 +0000 | [diff] [blame] | 1489 | while( N>x ){ |
drh | b37df7b | 2005-10-13 02:09:49 +0000 | [diff] [blame] | 1490 | logN += 1; |
drh | 28c4cf4 | 2005-07-27 20:41:43 +0000 | [diff] [blame] | 1491 | x *= 10; |
| 1492 | } |
| 1493 | return logN; |
| 1494 | } |
| 1495 | |
drh | 6d209d8 | 2006-06-27 01:54:26 +0000 | [diff] [blame] | 1496 | /* |
| 1497 | ** Two routines for printing the content of an sqlite3_index_info |
| 1498 | ** structure. Used for testing and debugging only. If neither |
| 1499 | ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines |
| 1500 | ** are no-ops. |
| 1501 | */ |
drh | 77a2a5e | 2007-04-06 01:04:39 +0000 | [diff] [blame] | 1502 | #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(SQLITE_DEBUG) |
drh | 6d209d8 | 2006-06-27 01:54:26 +0000 | [diff] [blame] | 1503 | static void TRACE_IDX_INPUTS(sqlite3_index_info *p){ |
| 1504 | int i; |
mlcreech | 3a00f90 | 2008-03-04 17:45:01 +0000 | [diff] [blame] | 1505 | if( !sqlite3WhereTrace ) return; |
drh | 6d209d8 | 2006-06-27 01:54:26 +0000 | [diff] [blame] | 1506 | for(i=0; i<p->nConstraint; i++){ |
| 1507 | sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", |
| 1508 | i, |
| 1509 | p->aConstraint[i].iColumn, |
| 1510 | p->aConstraint[i].iTermOffset, |
| 1511 | p->aConstraint[i].op, |
| 1512 | p->aConstraint[i].usable); |
| 1513 | } |
| 1514 | for(i=0; i<p->nOrderBy; i++){ |
| 1515 | sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", |
| 1516 | i, |
| 1517 | p->aOrderBy[i].iColumn, |
| 1518 | p->aOrderBy[i].desc); |
| 1519 | } |
| 1520 | } |
| 1521 | static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){ |
| 1522 | int i; |
mlcreech | 3a00f90 | 2008-03-04 17:45:01 +0000 | [diff] [blame] | 1523 | if( !sqlite3WhereTrace ) return; |
drh | 6d209d8 | 2006-06-27 01:54:26 +0000 | [diff] [blame] | 1524 | for(i=0; i<p->nConstraint; i++){ |
| 1525 | sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", |
| 1526 | i, |
| 1527 | p->aConstraintUsage[i].argvIndex, |
| 1528 | p->aConstraintUsage[i].omit); |
| 1529 | } |
| 1530 | sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); |
| 1531 | sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); |
| 1532 | sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); |
| 1533 | sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); |
| 1534 | } |
| 1535 | #else |
| 1536 | #define TRACE_IDX_INPUTS(A) |
| 1537 | #define TRACE_IDX_OUTPUTS(A) |
| 1538 | #endif |
| 1539 | |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1540 | /* |
| 1541 | ** Required because bestIndex() is called by bestOrClauseIndex() |
| 1542 | */ |
| 1543 | static void bestIndex( |
| 1544 | Parse*, WhereClause*, struct SrcList_item*, Bitmask, ExprList*, WhereCost*); |
| 1545 | |
| 1546 | /* |
| 1547 | ** This routine attempts to find an scanning strategy that can be used |
| 1548 | ** to optimize an 'OR' expression that is part of a WHERE clause. |
| 1549 | ** |
| 1550 | ** The table associated with FROM clause term pSrc may be either a |
| 1551 | ** regular B-Tree table or a virtual table. |
| 1552 | */ |
| 1553 | static void bestOrClauseIndex( |
| 1554 | Parse *pParse, /* The parsing context */ |
| 1555 | WhereClause *pWC, /* The WHERE clause */ |
| 1556 | struct SrcList_item *pSrc, /* The FROM clause term to search */ |
| 1557 | Bitmask notReady, /* Mask of cursors that are not available */ |
| 1558 | ExprList *pOrderBy, /* The ORDER BY clause */ |
| 1559 | WhereCost *pCost /* Lowest cost query plan */ |
| 1560 | ){ |
| 1561 | #ifndef SQLITE_OMIT_OR_OPTIMIZATION |
| 1562 | const int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */ |
| 1563 | const Bitmask maskSrc = getMask(pWC->pMaskSet, iCur); /* Bitmask for pSrc */ |
| 1564 | WhereTerm * const pWCEnd = &pWC->a[pWC->nTerm]; /* End of pWC->a[] */ |
| 1565 | WhereTerm *pTerm; /* A single term of the WHERE clause */ |
| 1566 | |
| 1567 | /* Search the WHERE clause terms for a usable WO_OR term. */ |
| 1568 | for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ |
| 1569 | if( pTerm->eOperator==WO_OR |
| 1570 | && ((pTerm->prereqAll & ~maskSrc) & notReady)==0 |
| 1571 | && (pTerm->u.pOrInfo->indexable & maskSrc)!=0 |
| 1572 | ){ |
| 1573 | WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; |
| 1574 | WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; |
| 1575 | WhereTerm *pOrTerm; |
| 1576 | int flags = WHERE_MULTI_OR; |
| 1577 | double rTotal = 0; |
| 1578 | double nRow = 0; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 1579 | Bitmask used = 0; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1580 | |
| 1581 | for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ |
| 1582 | WhereCost sTermCost; |
| 1583 | WHERETRACE(("... Multi-index OR testing for term %d of %d....\n", |
| 1584 | (pOrTerm - pOrWC->a), (pTerm - pWC->a) |
| 1585 | )); |
| 1586 | if( pOrTerm->eOperator==WO_AND ){ |
| 1587 | WhereClause *pAndWC = &pOrTerm->u.pAndInfo->wc; |
| 1588 | bestIndex(pParse, pAndWC, pSrc, notReady, 0, &sTermCost); |
| 1589 | }else if( pOrTerm->leftCursor==iCur ){ |
| 1590 | WhereClause tempWC; |
| 1591 | tempWC.pParse = pWC->pParse; |
| 1592 | tempWC.pMaskSet = pWC->pMaskSet; |
| 1593 | tempWC.op = TK_AND; |
| 1594 | tempWC.a = pOrTerm; |
| 1595 | tempWC.nTerm = 1; |
| 1596 | bestIndex(pParse, &tempWC, pSrc, notReady, 0, &sTermCost); |
| 1597 | }else{ |
| 1598 | continue; |
| 1599 | } |
| 1600 | rTotal += sTermCost.rCost; |
| 1601 | nRow += sTermCost.nRow; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 1602 | used |= sTermCost.used; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1603 | if( rTotal>=pCost->rCost ) break; |
| 1604 | } |
| 1605 | |
| 1606 | /* If there is an ORDER BY clause, increase the scan cost to account |
| 1607 | ** for the cost of the sort. */ |
| 1608 | if( pOrderBy!=0 ){ |
| 1609 | rTotal += nRow*estLog(nRow); |
| 1610 | WHERETRACE(("... sorting increases OR cost to %.9g\n", rTotal)); |
| 1611 | } |
| 1612 | |
| 1613 | /* If the cost of scanning using this OR term for optimization is |
| 1614 | ** less than the current cost stored in pCost, replace the contents |
| 1615 | ** of pCost. */ |
| 1616 | WHERETRACE(("... multi-index OR cost=%.9g nrow=%.9g\n", rTotal, nRow)); |
| 1617 | if( rTotal<pCost->rCost ){ |
| 1618 | pCost->rCost = rTotal; |
| 1619 | pCost->nRow = nRow; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 1620 | pCost->used = used; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1621 | pCost->plan.wsFlags = flags; |
| 1622 | pCost->plan.u.pTerm = pTerm; |
| 1623 | } |
| 1624 | } |
| 1625 | } |
| 1626 | #endif /* SQLITE_OMIT_OR_OPTIMIZATION */ |
| 1627 | } |
| 1628 | |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1629 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
| 1630 | /* |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1631 | ** Allocate and populate an sqlite3_index_info structure. It is the |
| 1632 | ** responsibility of the caller to eventually release the structure |
| 1633 | ** by passing the pointer returned by this function to sqlite3_free(). |
| 1634 | */ |
| 1635 | static sqlite3_index_info *allocateIndexInfo( |
| 1636 | Parse *pParse, |
| 1637 | WhereClause *pWC, |
| 1638 | struct SrcList_item *pSrc, |
| 1639 | ExprList *pOrderBy |
| 1640 | ){ |
| 1641 | int i, j; |
| 1642 | int nTerm; |
| 1643 | struct sqlite3_index_constraint *pIdxCons; |
| 1644 | struct sqlite3_index_orderby *pIdxOrderBy; |
| 1645 | struct sqlite3_index_constraint_usage *pUsage; |
| 1646 | WhereTerm *pTerm; |
| 1647 | int nOrderBy; |
| 1648 | sqlite3_index_info *pIdxInfo; |
| 1649 | |
| 1650 | WHERETRACE(("Recomputing index info for %s...\n", pSrc->pTab->zName)); |
| 1651 | |
| 1652 | /* Count the number of possible WHERE clause constraints referring |
| 1653 | ** to this virtual table */ |
| 1654 | for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ |
| 1655 | if( pTerm->leftCursor != pSrc->iCursor ) continue; |
| 1656 | assert( (pTerm->eOperator&(pTerm->eOperator-1))==0 ); |
| 1657 | testcase( pTerm->eOperator==WO_IN ); |
| 1658 | testcase( pTerm->eOperator==WO_ISNULL ); |
| 1659 | if( pTerm->eOperator & (WO_IN|WO_ISNULL) ) continue; |
| 1660 | nTerm++; |
| 1661 | } |
| 1662 | |
| 1663 | /* If the ORDER BY clause contains only columns in the current |
| 1664 | ** virtual table then allocate space for the aOrderBy part of |
| 1665 | ** the sqlite3_index_info structure. |
| 1666 | */ |
| 1667 | nOrderBy = 0; |
| 1668 | if( pOrderBy ){ |
| 1669 | for(i=0; i<pOrderBy->nExpr; i++){ |
| 1670 | Expr *pExpr = pOrderBy->a[i].pExpr; |
| 1671 | if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; |
| 1672 | } |
| 1673 | if( i==pOrderBy->nExpr ){ |
| 1674 | nOrderBy = pOrderBy->nExpr; |
| 1675 | } |
| 1676 | } |
| 1677 | |
| 1678 | /* Allocate the sqlite3_index_info structure |
| 1679 | */ |
| 1680 | pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) |
| 1681 | + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm |
| 1682 | + sizeof(*pIdxOrderBy)*nOrderBy ); |
| 1683 | if( pIdxInfo==0 ){ |
| 1684 | sqlite3ErrorMsg(pParse, "out of memory"); |
| 1685 | /* (double)0 In case of SQLITE_OMIT_FLOATING_POINT... */ |
| 1686 | return 0; |
| 1687 | } |
| 1688 | |
| 1689 | /* Initialize the structure. The sqlite3_index_info structure contains |
| 1690 | ** many fields that are declared "const" to prevent xBestIndex from |
| 1691 | ** changing them. We have to do some funky casting in order to |
| 1692 | ** initialize those fields. |
| 1693 | */ |
| 1694 | pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1]; |
| 1695 | pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; |
| 1696 | pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; |
| 1697 | *(int*)&pIdxInfo->nConstraint = nTerm; |
| 1698 | *(int*)&pIdxInfo->nOrderBy = nOrderBy; |
| 1699 | *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons; |
| 1700 | *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy; |
| 1701 | *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage = |
| 1702 | pUsage; |
| 1703 | |
| 1704 | for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ |
| 1705 | if( pTerm->leftCursor != pSrc->iCursor ) continue; |
| 1706 | assert( (pTerm->eOperator&(pTerm->eOperator-1))==0 ); |
| 1707 | testcase( pTerm->eOperator==WO_IN ); |
| 1708 | testcase( pTerm->eOperator==WO_ISNULL ); |
| 1709 | if( pTerm->eOperator & (WO_IN|WO_ISNULL) ) continue; |
| 1710 | pIdxCons[j].iColumn = pTerm->u.leftColumn; |
| 1711 | pIdxCons[j].iTermOffset = i; |
| 1712 | pIdxCons[j].op = (u8)pTerm->eOperator; |
| 1713 | /* The direct assignment in the previous line is possible only because |
| 1714 | ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The |
| 1715 | ** following asserts verify this fact. */ |
| 1716 | assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); |
| 1717 | assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); |
| 1718 | assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); |
| 1719 | assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); |
| 1720 | assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); |
| 1721 | assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH ); |
| 1722 | assert( pTerm->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) ); |
| 1723 | j++; |
| 1724 | } |
| 1725 | for(i=0; i<nOrderBy; i++){ |
| 1726 | Expr *pExpr = pOrderBy->a[i].pExpr; |
| 1727 | pIdxOrderBy[i].iColumn = pExpr->iColumn; |
| 1728 | pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder; |
| 1729 | } |
| 1730 | |
| 1731 | return pIdxInfo; |
| 1732 | } |
| 1733 | |
| 1734 | /* |
| 1735 | ** The table object reference passed as the second argument to this function |
| 1736 | ** must represent a virtual table. This function invokes the xBestIndex() |
| 1737 | ** method of the virtual table with the sqlite3_index_info pointer passed |
| 1738 | ** as the argument. |
| 1739 | ** |
| 1740 | ** If an error occurs, pParse is populated with an error message and a |
| 1741 | ** non-zero value is returned. Otherwise, 0 is returned and the output |
| 1742 | ** part of the sqlite3_index_info structure is left populated. |
| 1743 | ** |
| 1744 | ** Whether or not an error is returned, it is the responsibility of the |
| 1745 | ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates |
| 1746 | ** that this is required. |
| 1747 | */ |
| 1748 | static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ |
danielk1977 | 595a523 | 2009-07-24 17:58:53 +0000 | [diff] [blame] | 1749 | sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1750 | int i; |
| 1751 | int rc; |
| 1752 | |
| 1753 | (void)sqlite3SafetyOff(pParse->db); |
| 1754 | WHERETRACE(("xBestIndex for %s\n", pTab->zName)); |
| 1755 | TRACE_IDX_INPUTS(p); |
| 1756 | rc = pVtab->pModule->xBestIndex(pVtab, p); |
| 1757 | TRACE_IDX_OUTPUTS(p); |
| 1758 | (void)sqlite3SafetyOn(pParse->db); |
| 1759 | |
| 1760 | if( rc!=SQLITE_OK ){ |
| 1761 | if( rc==SQLITE_NOMEM ){ |
| 1762 | pParse->db->mallocFailed = 1; |
| 1763 | }else if( !pVtab->zErrMsg ){ |
| 1764 | sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); |
| 1765 | }else{ |
| 1766 | sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); |
| 1767 | } |
| 1768 | } |
| 1769 | sqlite3DbFree(pParse->db, pVtab->zErrMsg); |
| 1770 | pVtab->zErrMsg = 0; |
| 1771 | |
| 1772 | for(i=0; i<p->nConstraint; i++){ |
| 1773 | if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){ |
| 1774 | sqlite3ErrorMsg(pParse, |
| 1775 | "table %s: xBestIndex returned an invalid plan", pTab->zName); |
| 1776 | } |
| 1777 | } |
| 1778 | |
| 1779 | return pParse->nErr; |
| 1780 | } |
| 1781 | |
| 1782 | |
| 1783 | /* |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1784 | ** Compute the best index for a virtual table. |
| 1785 | ** |
| 1786 | ** The best index is computed by the xBestIndex method of the virtual |
| 1787 | ** table module. This routine is really just a wrapper that sets up |
| 1788 | ** the sqlite3_index_info structure that is used to communicate with |
| 1789 | ** xBestIndex. |
| 1790 | ** |
| 1791 | ** In a join, this routine might be called multiple times for the |
| 1792 | ** same virtual table. The sqlite3_index_info structure is created |
| 1793 | ** and initialized on the first invocation and reused on all subsequent |
| 1794 | ** invocations. The sqlite3_index_info structure is also used when |
| 1795 | ** code is generated to access the virtual table. The whereInfoDelete() |
| 1796 | ** routine takes care of freeing the sqlite3_index_info structure after |
| 1797 | ** everybody has finished with it. |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1798 | */ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1799 | static void bestVirtualIndex( |
| 1800 | Parse *pParse, /* The parsing context */ |
| 1801 | WhereClause *pWC, /* The WHERE clause */ |
| 1802 | struct SrcList_item *pSrc, /* The FROM clause term to search */ |
| 1803 | Bitmask notReady, /* Mask of cursors that are not available */ |
| 1804 | ExprList *pOrderBy, /* The order by clause */ |
| 1805 | WhereCost *pCost, /* Lowest cost query plan */ |
| 1806 | sqlite3_index_info **ppIdxInfo /* Index information passed to xBestIndex */ |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1807 | ){ |
| 1808 | Table *pTab = pSrc->pTab; |
| 1809 | sqlite3_index_info *pIdxInfo; |
| 1810 | struct sqlite3_index_constraint *pIdxCons; |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1811 | struct sqlite3_index_constraint_usage *pUsage; |
| 1812 | WhereTerm *pTerm; |
| 1813 | int i, j; |
| 1814 | int nOrderBy; |
| 1815 | |
danielk1977 | 6eacd28 | 2009-04-29 11:50:53 +0000 | [diff] [blame] | 1816 | /* Make sure wsFlags is initialized to some sane value. Otherwise, if the |
| 1817 | ** malloc in allocateIndexInfo() fails and this function returns leaving |
| 1818 | ** wsFlags in an uninitialized state, the caller may behave unpredictably. |
| 1819 | */ |
drh | 6a863cd | 2009-05-06 18:42:21 +0000 | [diff] [blame] | 1820 | memset(pCost, 0, sizeof(*pCost)); |
danielk1977 | 6eacd28 | 2009-04-29 11:50:53 +0000 | [diff] [blame] | 1821 | pCost->plan.wsFlags = WHERE_VIRTUALTABLE; |
| 1822 | |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1823 | /* If the sqlite3_index_info structure has not been previously |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1824 | ** allocated and initialized, then allocate and initialize it now. |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1825 | */ |
| 1826 | pIdxInfo = *ppIdxInfo; |
| 1827 | if( pIdxInfo==0 ){ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1828 | *ppIdxInfo = pIdxInfo = allocateIndexInfo(pParse, pWC, pSrc, pOrderBy); |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1829 | } |
danielk1977 | 732dc55 | 2009-04-21 17:23:04 +0000 | [diff] [blame] | 1830 | if( pIdxInfo==0 ){ |
| 1831 | return; |
| 1832 | } |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1833 | |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1834 | /* At this point, the sqlite3_index_info structure that pIdxInfo points |
| 1835 | ** to will have been initialized, either during the current invocation or |
| 1836 | ** during some prior invocation. Now we just have to customize the |
| 1837 | ** details of pIdxInfo for the current invocation and pass it to |
| 1838 | ** xBestIndex. |
| 1839 | */ |
| 1840 | |
danielk1977 | 935ed5e | 2007-03-30 09:13:13 +0000 | [diff] [blame] | 1841 | /* The module name must be defined. Also, by this point there must |
| 1842 | ** be a pointer to an sqlite3_vtab structure. Otherwise |
| 1843 | ** sqlite3ViewGetColumnNames() would have picked up the error. |
| 1844 | */ |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1845 | assert( pTab->azModuleArg && pTab->azModuleArg[0] ); |
danielk1977 | 595a523 | 2009-07-24 17:58:53 +0000 | [diff] [blame] | 1846 | assert( sqlite3GetVTable(pParse->db, pTab) ); |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1847 | |
| 1848 | /* Set the aConstraint[].usable fields and initialize all |
drh | 7f37590 | 2006-06-13 17:38:59 +0000 | [diff] [blame] | 1849 | ** output variables to zero. |
| 1850 | ** |
| 1851 | ** aConstraint[].usable is true for constraints where the right-hand |
| 1852 | ** side contains only references to tables to the left of the current |
| 1853 | ** table. In other words, if the constraint is of the form: |
| 1854 | ** |
| 1855 | ** column = expr |
| 1856 | ** |
| 1857 | ** and we are evaluating a join, then the constraint on column is |
| 1858 | ** only valid if all tables referenced in expr occur to the left |
| 1859 | ** of the table containing column. |
| 1860 | ** |
| 1861 | ** The aConstraints[] array contains entries for all constraints |
| 1862 | ** on the current table. That way we only have to compute it once |
| 1863 | ** even though we might try to pick the best index multiple times. |
| 1864 | ** For each attempt at picking an index, the order of tables in the |
| 1865 | ** join might be different so we have to recompute the usable flag |
| 1866 | ** each time. |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1867 | */ |
| 1868 | pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; |
| 1869 | pUsage = pIdxInfo->aConstraintUsage; |
| 1870 | for(i=0; i<pIdxInfo->nConstraint; i++, pIdxCons++){ |
| 1871 | j = pIdxCons->iTermOffset; |
| 1872 | pTerm = &pWC->a[j]; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 1873 | pIdxCons->usable = (pTerm->prereqRight¬Ready) ? 0 : 1; |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1874 | } |
| 1875 | memset(pUsage, 0, sizeof(pUsage[0])*pIdxInfo->nConstraint); |
drh | 4be8b51 | 2006-06-13 23:51:34 +0000 | [diff] [blame] | 1876 | if( pIdxInfo->needToFreeIdxStr ){ |
| 1877 | sqlite3_free(pIdxInfo->idxStr); |
| 1878 | } |
| 1879 | pIdxInfo->idxStr = 0; |
| 1880 | pIdxInfo->idxNum = 0; |
| 1881 | pIdxInfo->needToFreeIdxStr = 0; |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1882 | pIdxInfo->orderByConsumed = 0; |
shane | fbd60f8 | 2009-02-04 03:59:25 +0000 | [diff] [blame] | 1883 | /* ((double)2) In case of SQLITE_OMIT_FLOATING_POINT... */ |
| 1884 | pIdxInfo->estimatedCost = SQLITE_BIG_DBL / ((double)2); |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1885 | nOrderBy = pIdxInfo->nOrderBy; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1886 | if( !pOrderBy ){ |
| 1887 | pIdxInfo->nOrderBy = 0; |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1888 | } |
danielk1977 | 74cdba4 | 2006-06-19 12:02:58 +0000 | [diff] [blame] | 1889 | |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1890 | if( vtabBestIndex(pParse, pTab, pIdxInfo) ){ |
| 1891 | return; |
danielk1977 | 39359dc | 2008-03-17 09:36:44 +0000 | [diff] [blame] | 1892 | } |
| 1893 | |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 1894 | pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; |
| 1895 | for(i=0; i<pIdxInfo->nConstraint; i++){ |
| 1896 | if( pUsage[i].argvIndex>0 ){ |
| 1897 | pCost->used |= pWC->a[pIdxCons[i].iTermOffset].prereqRight; |
| 1898 | } |
| 1899 | } |
| 1900 | |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1901 | /* The cost is not allowed to be larger than SQLITE_BIG_DBL (the |
| 1902 | ** inital value of lowestCost in this loop. If it is, then the |
| 1903 | ** (cost<lowestCost) test below will never be true. |
| 1904 | ** |
| 1905 | ** Use "(double)2" instead of "2.0" in case OMIT_FLOATING_POINT |
| 1906 | ** is defined. |
| 1907 | */ |
| 1908 | if( (SQLITE_BIG_DBL/((double)2))<pIdxInfo->estimatedCost ){ |
| 1909 | pCost->rCost = (SQLITE_BIG_DBL/((double)2)); |
| 1910 | }else{ |
| 1911 | pCost->rCost = pIdxInfo->estimatedCost; |
| 1912 | } |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1913 | pCost->plan.u.pVtabIdx = pIdxInfo; |
drh | 5901b57 | 2009-06-10 19:33:28 +0000 | [diff] [blame] | 1914 | if( pIdxInfo->orderByConsumed ){ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 1915 | pCost->plan.wsFlags |= WHERE_ORDERBY; |
| 1916 | } |
| 1917 | pCost->plan.nEq = 0; |
| 1918 | pIdxInfo->nOrderBy = nOrderBy; |
| 1919 | |
| 1920 | /* Try to find a more efficient access pattern by using multiple indexes |
| 1921 | ** to optimize an OR expression within the WHERE clause. |
| 1922 | */ |
| 1923 | bestOrClauseIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost); |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 1924 | } |
| 1925 | #endif /* SQLITE_OMIT_VIRTUALTABLE */ |
| 1926 | |
drh | 28c4cf4 | 2005-07-27 20:41:43 +0000 | [diff] [blame] | 1927 | /* |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1928 | ** Argument pIdx is a pointer to an index structure that has an array of |
| 1929 | ** SQLITE_INDEX_SAMPLES evenly spaced samples of the first indexed column |
| 1930 | ** stored in Index.aSample. The domain of values stored in said column |
| 1931 | ** may be thought of as divided into (SQLITE_INDEX_SAMPLES+1) regions. |
| 1932 | ** Region 0 contains all values smaller than the first sample value. Region |
| 1933 | ** 1 contains values larger than or equal to the value of the first sample, |
| 1934 | ** but smaller than the value of the second. And so on. |
| 1935 | ** |
| 1936 | ** If successful, this function determines which of the regions value |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 1937 | ** pVal lies in, sets *piRegion to the region index (a value between 0 |
| 1938 | ** and SQLITE_INDEX_SAMPLES+1, inclusive) and returns SQLITE_OK. |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1939 | ** Or, if an OOM occurs while converting text values between encodings, |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 1940 | ** SQLITE_NOMEM is returned and *piRegion is undefined. |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1941 | */ |
dan | 69188d9 | 2009-08-19 08:18:32 +0000 | [diff] [blame] | 1942 | #ifdef SQLITE_ENABLE_STAT2 |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1943 | static int whereRangeRegion( |
| 1944 | Parse *pParse, /* Database connection */ |
| 1945 | Index *pIdx, /* Index to consider domain of */ |
| 1946 | sqlite3_value *pVal, /* Value to consider */ |
| 1947 | int *piRegion /* OUT: Region of domain in which value lies */ |
| 1948 | ){ |
drh | daf4a9f | 2009-08-20 20:05:55 +0000 | [diff] [blame] | 1949 | if( ALWAYS(pVal) ){ |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1950 | IndexSample *aSample = pIdx->aSample; |
| 1951 | int i = 0; |
| 1952 | int eType = sqlite3_value_type(pVal); |
| 1953 | |
| 1954 | if( eType==SQLITE_INTEGER || eType==SQLITE_FLOAT ){ |
| 1955 | double r = sqlite3_value_double(pVal); |
| 1956 | for(i=0; i<SQLITE_INDEX_SAMPLES; i++){ |
| 1957 | if( aSample[i].eType==SQLITE_NULL ) continue; |
| 1958 | if( aSample[i].eType>=SQLITE_TEXT || aSample[i].u.r>r ) break; |
| 1959 | } |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 1960 | }else{ |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1961 | sqlite3 *db = pParse->db; |
| 1962 | CollSeq *pColl; |
| 1963 | const u8 *z; |
| 1964 | int n; |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 1965 | |
| 1966 | /* pVal comes from sqlite3ValueFromExpr() so the type cannot be NULL */ |
| 1967 | assert( eType==SQLITE_TEXT || eType==SQLITE_BLOB ); |
| 1968 | |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1969 | if( eType==SQLITE_BLOB ){ |
| 1970 | z = (const u8 *)sqlite3_value_blob(pVal); |
| 1971 | pColl = db->pDfltColl; |
dan | e275dc3 | 2009-08-18 16:24:58 +0000 | [diff] [blame] | 1972 | assert( pColl->enc==SQLITE_UTF8 ); |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1973 | }else{ |
drh | 9aeda79 | 2009-08-20 02:34:15 +0000 | [diff] [blame] | 1974 | pColl = sqlite3GetCollSeq(db, SQLITE_UTF8, 0, *pIdx->azColl); |
| 1975 | if( pColl==0 ){ |
| 1976 | sqlite3ErrorMsg(pParse, "no such collation sequence: %s", |
| 1977 | *pIdx->azColl); |
dan | e275dc3 | 2009-08-18 16:24:58 +0000 | [diff] [blame] | 1978 | return SQLITE_ERROR; |
| 1979 | } |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1980 | z = (const u8 *)sqlite3ValueText(pVal, pColl->enc); |
dan | e275dc3 | 2009-08-18 16:24:58 +0000 | [diff] [blame] | 1981 | if( !z ){ |
| 1982 | return SQLITE_NOMEM; |
| 1983 | } |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1984 | assert( z && pColl && pColl->xCmp ); |
| 1985 | } |
| 1986 | n = sqlite3ValueBytes(pVal, pColl->enc); |
| 1987 | |
| 1988 | for(i=0; i<SQLITE_INDEX_SAMPLES; i++){ |
dan | e275dc3 | 2009-08-18 16:24:58 +0000 | [diff] [blame] | 1989 | int r; |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1990 | int eSampletype = aSample[i].eType; |
| 1991 | if( eSampletype==SQLITE_NULL || eSampletype<eType ) continue; |
| 1992 | if( (eSampletype!=eType) ) break; |
dan | e83c4f3 | 2009-09-21 16:34:24 +0000 | [diff] [blame] | 1993 | #ifndef SQLITE_OMIT_UTF16 |
| 1994 | if( pColl->enc!=SQLITE_UTF8 ){ |
dan | e275dc3 | 2009-08-18 16:24:58 +0000 | [diff] [blame] | 1995 | int nSample; |
| 1996 | char *zSample = sqlite3Utf8to16( |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 1997 | db, pColl->enc, aSample[i].u.z, aSample[i].nByte, &nSample |
| 1998 | ); |
dan | e275dc3 | 2009-08-18 16:24:58 +0000 | [diff] [blame] | 1999 | if( !zSample ){ |
| 2000 | assert( db->mallocFailed ); |
| 2001 | return SQLITE_NOMEM; |
| 2002 | } |
| 2003 | r = pColl->xCmp(pColl->pUser, nSample, zSample, n, z); |
| 2004 | sqlite3DbFree(db, zSample); |
dan | e83c4f3 | 2009-09-21 16:34:24 +0000 | [diff] [blame] | 2005 | }else |
| 2006 | #endif |
| 2007 | { |
| 2008 | r = pColl->xCmp(pColl->pUser, aSample[i].nByte, aSample[i].u.z, n, z); |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2009 | } |
dan | e275dc3 | 2009-08-18 16:24:58 +0000 | [diff] [blame] | 2010 | if( r>0 ) break; |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2011 | } |
| 2012 | } |
| 2013 | |
drh | a8f5761 | 2009-08-25 16:28:14 +0000 | [diff] [blame] | 2014 | assert( i>=0 && i<=SQLITE_INDEX_SAMPLES ); |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2015 | *piRegion = i; |
| 2016 | } |
| 2017 | return SQLITE_OK; |
| 2018 | } |
dan | 69188d9 | 2009-08-19 08:18:32 +0000 | [diff] [blame] | 2019 | #endif /* #ifdef SQLITE_ENABLE_STAT2 */ |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2020 | |
| 2021 | /* |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2022 | ** If expression pExpr represents a literal value, set *pp to point to |
| 2023 | ** an sqlite3_value structure containing the same value, with affinity |
| 2024 | ** aff applied to it, before returning. It is the responsibility of the |
| 2025 | ** caller to eventually release this structure by passing it to |
| 2026 | ** sqlite3ValueFree(). |
| 2027 | ** |
| 2028 | ** If the current parse is a recompile (sqlite3Reprepare()) and pExpr |
| 2029 | ** is an SQL variable that currently has a non-NULL value bound to it, |
| 2030 | ** create an sqlite3_value structure containing this value, again with |
| 2031 | ** affinity aff applied to it, instead. |
| 2032 | ** |
| 2033 | ** If neither of the above apply, set *pp to NULL. |
| 2034 | ** |
| 2035 | ** If an error occurs, return an error code. Otherwise, SQLITE_OK. |
| 2036 | */ |
dan | f7b0b0a | 2009-10-19 15:52:32 +0000 | [diff] [blame] | 2037 | #ifdef SQLITE_ENABLE_STAT2 |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2038 | static int valueFromExpr( |
| 2039 | Parse *pParse, |
| 2040 | Expr *pExpr, |
| 2041 | u8 aff, |
| 2042 | sqlite3_value **pp |
| 2043 | ){ |
drh | b4138de | 2009-10-19 22:41:06 +0000 | [diff] [blame] | 2044 | /* The evalConstExpr() function will have already converted any TK_VARIABLE |
| 2045 | ** expression involved in an comparison into a TK_REGISTER. */ |
| 2046 | assert( pExpr->op!=TK_VARIABLE ); |
| 2047 | if( pExpr->op==TK_REGISTER && pExpr->op2==TK_VARIABLE ){ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2048 | int iVar = pExpr->iColumn; |
dan | 1d2ce4f | 2009-10-19 18:11:09 +0000 | [diff] [blame] | 2049 | sqlite3VdbeSetVarmask(pParse->pVdbe, iVar); |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2050 | *pp = sqlite3VdbeGetValue(pParse->pReprepare, iVar, aff); |
| 2051 | return SQLITE_OK; |
| 2052 | } |
| 2053 | return sqlite3ValueFromExpr(pParse->db, pExpr, SQLITE_UTF8, aff, pp); |
| 2054 | } |
dan | f7b0b0a | 2009-10-19 15:52:32 +0000 | [diff] [blame] | 2055 | #endif |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2056 | |
| 2057 | /* |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2058 | ** This function is used to estimate the number of rows that will be visited |
| 2059 | ** by scanning an index for a range of values. The range may have an upper |
| 2060 | ** bound, a lower bound, or both. The WHERE clause terms that set the upper |
| 2061 | ** and lower bounds are represented by pLower and pUpper respectively. For |
| 2062 | ** example, assuming that index p is on t1(a): |
| 2063 | ** |
| 2064 | ** ... FROM t1 WHERE a > ? AND a < ? ... |
| 2065 | ** |_____| |_____| |
| 2066 | ** | | |
| 2067 | ** pLower pUpper |
| 2068 | ** |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2069 | ** If either of the upper or lower bound is not present, then NULL is passed in |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2070 | ** place of the corresponding WhereTerm. |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2071 | ** |
| 2072 | ** The nEq parameter is passed the index of the index column subject to the |
| 2073 | ** range constraint. Or, equivalently, the number of equality constraints |
| 2074 | ** optimized by the proposed index scan. For example, assuming index p is |
| 2075 | ** on t1(a, b), and the SQL query is: |
| 2076 | ** |
| 2077 | ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... |
| 2078 | ** |
| 2079 | ** then nEq should be passed the value 1 (as the range restricted column, |
| 2080 | ** b, is the second left-most column of the index). Or, if the query is: |
| 2081 | ** |
| 2082 | ** ... FROM t1 WHERE a > ? AND a < ? ... |
| 2083 | ** |
| 2084 | ** then nEq should be passed 0. |
| 2085 | ** |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2086 | ** The returned value is an integer between 1 and 100, inclusive. A return |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2087 | ** value of 1 indicates that the proposed range scan is expected to visit |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2088 | ** approximately 1/100th (1%) of the rows selected by the nEq equality |
| 2089 | ** constraints (if any). A return value of 100 indicates that it is expected |
| 2090 | ** that the range scan will visit every row (100%) selected by the equality |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2091 | ** constraints. |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2092 | ** |
| 2093 | ** In the absence of sqlite_stat2 ANALYZE data, each range inequality |
| 2094 | ** reduces the search space by 2/3rds. Hence a single constraint (x>?) |
| 2095 | ** results in a return of 33 and a range constraint (x>? AND x<?) results |
| 2096 | ** in a return of 11. |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2097 | */ |
| 2098 | static int whereRangeScanEst( |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2099 | Parse *pParse, /* Parsing & code generating context */ |
| 2100 | Index *p, /* The index containing the range-compared column; "x" */ |
| 2101 | int nEq, /* index into p->aCol[] of the range-compared column */ |
| 2102 | WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ |
| 2103 | WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ |
| 2104 | int *piEst /* OUT: Return value */ |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2105 | ){ |
dan | 69188d9 | 2009-08-19 08:18:32 +0000 | [diff] [blame] | 2106 | int rc = SQLITE_OK; |
| 2107 | |
| 2108 | #ifdef SQLITE_ENABLE_STAT2 |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2109 | |
| 2110 | if( nEq==0 && p->aSample ){ |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2111 | sqlite3_value *pLowerVal = 0; |
| 2112 | sqlite3_value *pUpperVal = 0; |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2113 | int iEst; |
drh | 011cfca | 2009-08-25 15:56:51 +0000 | [diff] [blame] | 2114 | int iLower = 0; |
| 2115 | int iUpper = SQLITE_INDEX_SAMPLES; |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2116 | u8 aff = p->pTable->aCol[p->aiColumn[0]].affinity; |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2117 | |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2118 | if( pLower ){ |
| 2119 | Expr *pExpr = pLower->pExpr->pRight; |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2120 | rc = valueFromExpr(pParse, pExpr, aff, &pLowerVal); |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2121 | } |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2122 | if( rc==SQLITE_OK && pUpper ){ |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2123 | Expr *pExpr = pUpper->pExpr->pRight; |
dan | 937d0de | 2009-10-15 18:35:38 +0000 | [diff] [blame] | 2124 | rc = valueFromExpr(pParse, pExpr, aff, &pUpperVal); |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2125 | } |
| 2126 | |
| 2127 | if( rc!=SQLITE_OK || (pLowerVal==0 && pUpperVal==0) ){ |
| 2128 | sqlite3ValueFree(pLowerVal); |
| 2129 | sqlite3ValueFree(pUpperVal); |
| 2130 | goto range_est_fallback; |
| 2131 | }else if( pLowerVal==0 ){ |
| 2132 | rc = whereRangeRegion(pParse, p, pUpperVal, &iUpper); |
drh | 011cfca | 2009-08-25 15:56:51 +0000 | [diff] [blame] | 2133 | if( pLower ) iLower = iUpper/2; |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2134 | }else if( pUpperVal==0 ){ |
| 2135 | rc = whereRangeRegion(pParse, p, pLowerVal, &iLower); |
drh | 011cfca | 2009-08-25 15:56:51 +0000 | [diff] [blame] | 2136 | if( pUpper ) iUpper = (iLower + SQLITE_INDEX_SAMPLES + 1)/2; |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2137 | }else{ |
| 2138 | rc = whereRangeRegion(pParse, p, pUpperVal, &iUpper); |
| 2139 | if( rc==SQLITE_OK ){ |
| 2140 | rc = whereRangeRegion(pParse, p, pLowerVal, &iLower); |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2141 | } |
| 2142 | } |
| 2143 | |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2144 | iEst = iUpper - iLower; |
drh | a8f5761 | 2009-08-25 16:28:14 +0000 | [diff] [blame] | 2145 | testcase( iEst==SQLITE_INDEX_SAMPLES ); |
| 2146 | assert( iEst<=SQLITE_INDEX_SAMPLES ); |
| 2147 | if( iEst<1 ){ |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2148 | iEst = 1; |
| 2149 | } |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2150 | |
| 2151 | sqlite3ValueFree(pLowerVal); |
| 2152 | sqlite3ValueFree(pUpperVal); |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2153 | *piEst = (iEst * 100)/SQLITE_INDEX_SAMPLES; |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2154 | return rc; |
| 2155 | } |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2156 | range_est_fallback: |
drh | 3f02218 | 2009-09-09 16:10:50 +0000 | [diff] [blame] | 2157 | #else |
| 2158 | UNUSED_PARAMETER(pParse); |
| 2159 | UNUSED_PARAMETER(p); |
| 2160 | UNUSED_PARAMETER(nEq); |
dan | 69188d9 | 2009-08-19 08:18:32 +0000 | [diff] [blame] | 2161 | #endif |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2162 | assert( pLower || pUpper ); |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2163 | if( pLower && pUpper ){ |
| 2164 | *piEst = 11; |
| 2165 | }else{ |
| 2166 | *piEst = 33; |
| 2167 | } |
dan | 02fa469 | 2009-08-17 17:06:58 +0000 | [diff] [blame] | 2168 | return rc; |
| 2169 | } |
| 2170 | |
| 2171 | |
| 2172 | /* |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2173 | ** Find the query plan for accessing a particular table. Write the |
| 2174 | ** best query plan and its cost into the WhereCost object supplied as the |
| 2175 | ** last parameter. |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2176 | ** |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2177 | ** The lowest cost plan wins. The cost is an estimate of the amount of |
| 2178 | ** CPU and disk I/O need to process the request using the selected plan. |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2179 | ** Factors that influence cost include: |
| 2180 | ** |
| 2181 | ** * The estimated number of rows that will be retrieved. (The |
| 2182 | ** fewer the better.) |
| 2183 | ** |
| 2184 | ** * Whether or not sorting must occur. |
| 2185 | ** |
| 2186 | ** * Whether or not there must be separate lookups in the |
| 2187 | ** index and in the main table. |
| 2188 | ** |
danielk1977 | e2d7b24 | 2009-02-23 17:33:49 +0000 | [diff] [blame] | 2189 | ** If there was an INDEXED BY clause (pSrc->pIndex) attached to the table in |
| 2190 | ** the SQL statement, then this function only considers plans using the |
drh | 296a483 | 2009-03-22 20:36:18 +0000 | [diff] [blame] | 2191 | ** named index. If no such plan is found, then the returned cost is |
| 2192 | ** SQLITE_BIG_DBL. If a plan is found that uses the named index, |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 2193 | ** then the cost is calculated in the usual way. |
| 2194 | ** |
danielk1977 | e2d7b24 | 2009-02-23 17:33:49 +0000 | [diff] [blame] | 2195 | ** If a NOT INDEXED clause (pSrc->notIndexed!=0) was attached to the table |
| 2196 | ** in the SELECT statement, then no indexes are considered. However, the |
| 2197 | ** selected plan may still take advantage of the tables built-in rowid |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 2198 | ** index. |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2199 | */ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2200 | static void bestBtreeIndex( |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2201 | Parse *pParse, /* The parsing context */ |
| 2202 | WhereClause *pWC, /* The WHERE clause */ |
| 2203 | struct SrcList_item *pSrc, /* The FROM clause term to search */ |
| 2204 | Bitmask notReady, /* Mask of cursors that are not available */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2205 | ExprList *pOrderBy, /* The ORDER BY clause */ |
| 2206 | WhereCost *pCost /* Lowest cost query plan */ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2207 | ){ |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2208 | int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */ |
| 2209 | Index *pProbe; /* An index we are evaluating */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2210 | Index *pIdx; /* Copy of pProbe, or zero for IPK index */ |
| 2211 | int eqTermMask; /* Current mask of valid equality operators */ |
| 2212 | int idxEqTermMask; /* Index mask of valid equality operators */ |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2213 | Index sPk; /* A fake index object for the primary key */ |
| 2214 | unsigned int aiRowEstPk[2]; /* The aiRowEst[] value for the sPk index */ |
| 2215 | int aiColumnPk = -1; /* The aColumn[] value for the sPk index */ |
| 2216 | int wsFlagMask; /* Allowed flags in pCost->plan.wsFlag */ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2217 | |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2218 | /* Initialize the cost to a worst-case value */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2219 | memset(pCost, 0, sizeof(*pCost)); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2220 | pCost->rCost = SQLITE_BIG_DBL; |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2221 | |
drh | c49de5d | 2007-01-19 01:06:01 +0000 | [diff] [blame] | 2222 | /* If the pSrc table is the right table of a LEFT JOIN then we may not |
| 2223 | ** use an index to satisfy IS NULL constraints on that table. This is |
| 2224 | ** because columns might end up being NULL if the table does not match - |
| 2225 | ** a circumstance which the index cannot help us discover. Ticket #2177. |
| 2226 | */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2227 | if( pSrc->jointype & JT_LEFT ){ |
| 2228 | idxEqTermMask = WO_EQ|WO_IN; |
drh | c49de5d | 2007-01-19 01:06:01 +0000 | [diff] [blame] | 2229 | }else{ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2230 | idxEqTermMask = WO_EQ|WO_IN|WO_ISNULL; |
drh | c49de5d | 2007-01-19 01:06:01 +0000 | [diff] [blame] | 2231 | } |
| 2232 | |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 2233 | if( pSrc->pIndex ){ |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2234 | /* An INDEXED BY clause specifies a particular index to use */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2235 | pIdx = pProbe = pSrc->pIndex; |
| 2236 | wsFlagMask = ~(WHERE_ROWID_EQ|WHERE_ROWID_RANGE); |
| 2237 | eqTermMask = idxEqTermMask; |
| 2238 | }else{ |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2239 | /* There is no INDEXED BY clause. Create a fake Index object to |
| 2240 | ** represent the primary key */ |
| 2241 | Index *pFirst; /* Any other index on the table */ |
| 2242 | memset(&sPk, 0, sizeof(Index)); |
| 2243 | sPk.nColumn = 1; |
| 2244 | sPk.aiColumn = &aiColumnPk; |
| 2245 | sPk.aiRowEst = aiRowEstPk; |
| 2246 | aiRowEstPk[1] = 1; |
| 2247 | sPk.onError = OE_Replace; |
| 2248 | sPk.pTable = pSrc->pTab; |
| 2249 | pFirst = pSrc->pTab->pIndex; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2250 | if( pSrc->notIndexed==0 ){ |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2251 | sPk.pNext = pFirst; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2252 | } |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2253 | /* The aiRowEstPk[0] is an estimate of the total number of rows in the |
| 2254 | ** table. Get this information from the ANALYZE information if it is |
| 2255 | ** available. If not available, assume the table 1 million rows in size. |
| 2256 | */ |
| 2257 | if( pFirst ){ |
| 2258 | assert( pFirst->aiRowEst!=0 ); /* Allocated together with pFirst */ |
| 2259 | aiRowEstPk[0] = pFirst->aiRowEst[0]; |
| 2260 | }else{ |
| 2261 | aiRowEstPk[0] = 1000000; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2262 | } |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2263 | pProbe = &sPk; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2264 | wsFlagMask = ~( |
| 2265 | WHERE_COLUMN_IN|WHERE_COLUMN_EQ|WHERE_COLUMN_NULL|WHERE_COLUMN_RANGE |
| 2266 | ); |
| 2267 | eqTermMask = WO_EQ|WO_IN; |
| 2268 | pIdx = 0; |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 2269 | } |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2270 | |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2271 | /* Loop over all indices looking for the best one to use |
| 2272 | */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2273 | for(; pProbe; pIdx=pProbe=pProbe->pNext){ |
| 2274 | const unsigned int * const aiRowEst = pProbe->aiRowEst; |
| 2275 | double cost; /* Cost of using pProbe */ |
| 2276 | double nRow; /* Estimated number of rows in result set */ |
| 2277 | int rev; /* True to scan in reverse order */ |
| 2278 | int wsFlags = 0; |
| 2279 | Bitmask used = 0; |
| 2280 | |
| 2281 | /* The following variables are populated based on the properties of |
| 2282 | ** scan being evaluated. They are then used to determine the expected |
| 2283 | ** cost and number of rows returned. |
| 2284 | ** |
| 2285 | ** nEq: |
| 2286 | ** Number of equality terms that can be implemented using the index. |
| 2287 | ** |
| 2288 | ** nInMul: |
| 2289 | ** The "in-multiplier". This is an estimate of how many seek operations |
| 2290 | ** SQLite must perform on the index in question. For example, if the |
| 2291 | ** WHERE clause is: |
| 2292 | ** |
| 2293 | ** WHERE a IN (1, 2, 3) AND b IN (4, 5, 6) |
| 2294 | ** |
| 2295 | ** SQLite must perform 9 lookups on an index on (a, b), so nInMul is |
| 2296 | ** set to 9. Given the same schema and either of the following WHERE |
| 2297 | ** clauses: |
| 2298 | ** |
| 2299 | ** WHERE a = 1 |
| 2300 | ** WHERE a >= 2 |
| 2301 | ** |
| 2302 | ** nInMul is set to 1. |
| 2303 | ** |
| 2304 | ** If there exists a WHERE term of the form "x IN (SELECT ...)", then |
| 2305 | ** the sub-select is assumed to return 25 rows for the purposes of |
| 2306 | ** determining nInMul. |
| 2307 | ** |
| 2308 | ** bInEst: |
| 2309 | ** Set to true if there was at least one "x IN (SELECT ...)" term used |
| 2310 | ** in determining the value of nInMul. |
| 2311 | ** |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2312 | ** nBound: |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2313 | ** An estimate on the amount of the table that must be searched. A |
| 2314 | ** value of 100 means the entire table is searched. Range constraints |
| 2315 | ** might reduce this to a value less than 100 to indicate that only |
| 2316 | ** a fraction of the table needs searching. In the absence of |
| 2317 | ** sqlite_stat2 ANALYZE data, a single inequality reduces the search |
| 2318 | ** space to 1/3rd its original size. So an x>? constraint reduces |
| 2319 | ** nBound to 33. Two constraints (x>? AND x<?) reduce nBound to 11. |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2320 | ** |
| 2321 | ** bSort: |
| 2322 | ** Boolean. True if there is an ORDER BY clause that will require an |
| 2323 | ** external sort (i.e. scanning the index being evaluated will not |
| 2324 | ** correctly order records). |
| 2325 | ** |
| 2326 | ** bLookup: |
| 2327 | ** Boolean. True if for each index entry visited a lookup on the |
| 2328 | ** corresponding table b-tree is required. This is always false |
| 2329 | ** for the rowid index. For other indexes, it is true unless all the |
| 2330 | ** columns of the table used by the SELECT statement are present in |
| 2331 | ** the index (such an index is sometimes described as a covering index). |
| 2332 | ** For example, given the index on (a, b), the second of the following |
| 2333 | ** two queries requires table b-tree lookups, but the first does not. |
| 2334 | ** |
| 2335 | ** SELECT a, b FROM tbl WHERE a = 1; |
| 2336 | ** SELECT a, b, c FROM tbl WHERE a = 1; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2337 | */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2338 | int nEq; |
| 2339 | int bInEst = 0; |
| 2340 | int nInMul = 1; |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2341 | int nBound = 100; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2342 | int bSort = 0; |
| 2343 | int bLookup = 0; |
| 2344 | |
| 2345 | /* Determine the values of nEq and nInMul */ |
| 2346 | for(nEq=0; nEq<pProbe->nColumn; nEq++){ |
| 2347 | WhereTerm *pTerm; /* A single term of the WHERE clause */ |
| 2348 | int j = pProbe->aiColumn[nEq]; |
| 2349 | pTerm = findTerm(pWC, iCur, j, notReady, eqTermMask, pIdx); |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2350 | if( pTerm==0 ) break; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2351 | wsFlags |= (WHERE_COLUMN_EQ|WHERE_ROWID_EQ); |
drh | b52076c | 2006-01-23 13:22:09 +0000 | [diff] [blame] | 2352 | if( pTerm->eOperator & WO_IN ){ |
drh | a611040 | 2005-07-28 20:51:19 +0000 | [diff] [blame] | 2353 | Expr *pExpr = pTerm->pExpr; |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 2354 | wsFlags |= WHERE_COLUMN_IN; |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 2355 | if( ExprHasProperty(pExpr, EP_xIsSelect) ){ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2356 | nInMul *= 25; |
| 2357 | bInEst = 1; |
danielk1977 | 6ab3a2e | 2009-02-19 14:39:25 +0000 | [diff] [blame] | 2358 | }else if( pExpr->x.pList ){ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2359 | nInMul *= pExpr->x.pList->nExpr + 1; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2360 | } |
drh | 46619d6 | 2009-04-24 14:51:42 +0000 | [diff] [blame] | 2361 | }else if( pTerm->eOperator & WO_ISNULL ){ |
| 2362 | wsFlags |= WHERE_COLUMN_NULL; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2363 | } |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2364 | used |= pTerm->prereqRight; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2365 | } |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2366 | |
| 2367 | /* Determine the value of nBound. */ |
| 2368 | if( nEq<pProbe->nColumn ){ |
| 2369 | int j = pProbe->aiColumn[nEq]; |
| 2370 | if( findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE|WO_GT|WO_GE, pIdx) ){ |
| 2371 | WhereTerm *pTop = findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE, pIdx); |
| 2372 | WhereTerm *pBtm = findTerm(pWC, iCur, j, notReady, WO_GT|WO_GE, pIdx); |
dan | e275dc3 | 2009-08-18 16:24:58 +0000 | [diff] [blame] | 2373 | whereRangeScanEst(pParse, pProbe, nEq, pBtm, pTop, &nBound); |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2374 | if( pTop ){ |
| 2375 | wsFlags |= WHERE_TOP_LIMIT; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2376 | used |= pTop->prereqRight; |
| 2377 | } |
| 2378 | if( pBtm ){ |
| 2379 | wsFlags |= WHERE_BTM_LIMIT; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2380 | used |= pBtm->prereqRight; |
| 2381 | } |
| 2382 | wsFlags |= (WHERE_COLUMN_RANGE|WHERE_ROWID_RANGE); |
| 2383 | } |
| 2384 | }else if( pProbe->onError!=OE_None ){ |
drh | 46619d6 | 2009-04-24 14:51:42 +0000 | [diff] [blame] | 2385 | testcase( wsFlags & WHERE_COLUMN_IN ); |
| 2386 | testcase( wsFlags & WHERE_COLUMN_NULL ); |
| 2387 | if( (wsFlags & (WHERE_COLUMN_IN|WHERE_COLUMN_NULL))==0 ){ |
| 2388 | wsFlags |= WHERE_UNIQUE; |
| 2389 | } |
drh | 943af3c | 2005-07-29 19:43:58 +0000 | [diff] [blame] | 2390 | } |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2391 | |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2392 | /* If there is an ORDER BY clause and the index being considered will |
| 2393 | ** naturally scan rows in the required order, set the appropriate flags |
| 2394 | ** in wsFlags. Otherwise, if there is an ORDER BY clause but the index |
| 2395 | ** will scan rows in a different order, set the bSort variable. */ |
drh | 28c4cf4 | 2005-07-27 20:41:43 +0000 | [diff] [blame] | 2396 | if( pOrderBy ){ |
drh | 46619d6 | 2009-04-24 14:51:42 +0000 | [diff] [blame] | 2397 | if( (wsFlags & (WHERE_COLUMN_IN|WHERE_COLUMN_NULL))==0 |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2398 | && isSortingIndex(pParse,pWC->pMaskSet,pProbe,iCur,pOrderBy,nEq,&rev) |
drh | 46619d6 | 2009-04-24 14:51:42 +0000 | [diff] [blame] | 2399 | ){ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2400 | wsFlags |= WHERE_ROWID_RANGE|WHERE_COLUMN_RANGE|WHERE_ORDERBY; |
| 2401 | wsFlags |= (rev ? WHERE_REVERSE : 0); |
drh | 28c4cf4 | 2005-07-27 20:41:43 +0000 | [diff] [blame] | 2402 | }else{ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2403 | bSort = 1; |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2404 | } |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2405 | } |
| 2406 | |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2407 | /* If currently calculating the cost of using an index (not the IPK |
| 2408 | ** index), determine if all required column data may be obtained without |
| 2409 | ** seeking to entries in the main table (i.e. if the index is a covering |
| 2410 | ** index for this query). If it is, set the WHERE_IDX_ONLY flag in |
| 2411 | ** wsFlags. Otherwise, set the bLookup variable to true. */ |
| 2412 | if( pIdx && wsFlags ){ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2413 | Bitmask m = pSrc->colUsed; |
| 2414 | int j; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2415 | for(j=0; j<pIdx->nColumn; j++){ |
| 2416 | int x = pIdx->aiColumn[j]; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2417 | if( x<BMS-1 ){ |
| 2418 | m &= ~(((Bitmask)1)<<x); |
| 2419 | } |
| 2420 | } |
| 2421 | if( m==0 ){ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 2422 | wsFlags |= WHERE_IDX_ONLY; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2423 | }else{ |
| 2424 | bLookup = 1; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2425 | } |
| 2426 | } |
| 2427 | |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2428 | /**** Begin adding up the cost of using this index (Needs improvements) |
| 2429 | ** |
| 2430 | ** Estimate the number of rows of output. For an IN operator, |
| 2431 | ** do not let the estimate exceed half the rows in the table. |
| 2432 | */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2433 | nRow = (double)(aiRowEst[nEq] * nInMul); |
| 2434 | if( bInEst && nRow*2>aiRowEst[0] ){ |
| 2435 | nRow = aiRowEst[0]/2; |
shane | cea72b2 | 2009-09-07 04:38:36 +0000 | [diff] [blame] | 2436 | nInMul = (int)(nRow / aiRowEst[nEq]); |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2437 | } |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2438 | |
| 2439 | /* Assume constant cost to access a row and logarithmic cost to |
| 2440 | ** do a binary search. Hence, the initial cost is the number of output |
| 2441 | ** rows plus log2(table-size) times the number of binary searches. |
| 2442 | */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2443 | cost = nRow + nInMul*estLog(aiRowEst[0]); |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2444 | |
| 2445 | /* Adjust the number of rows and the cost downward to reflect rows |
| 2446 | ** that are excluded by range constraints. |
| 2447 | */ |
drh | 98cdf62 | 2009-08-20 18:14:42 +0000 | [diff] [blame] | 2448 | nRow = (nRow * (double)nBound) / (double)100; |
| 2449 | cost = (cost * (double)nBound) / (double)100; |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2450 | |
| 2451 | /* Add in the estimated cost of sorting the result |
| 2452 | */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2453 | if( bSort ){ |
| 2454 | cost += cost*estLog(cost); |
| 2455 | } |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2456 | |
| 2457 | /* If all information can be taken directly from the index, we avoid |
| 2458 | ** doing table lookups. This reduces the cost by half. (Not really - |
| 2459 | ** this needs to be fixed.) |
| 2460 | */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2461 | if( pIdx && bLookup==0 ){ |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2462 | cost /= (double)2; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2463 | } |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2464 | /**** Cost of using this index has now been computed ****/ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2465 | |
| 2466 | WHERETRACE(( |
| 2467 | "tbl=%s idx=%s nEq=%d nInMul=%d nBound=%d bSort=%d bLookup=%d" |
| 2468 | " wsFlags=%d (nRow=%.2f cost=%.2f)\n", |
| 2469 | pSrc->pTab->zName, (pIdx ? pIdx->zName : "ipk"), |
| 2470 | nEq, nInMul, nBound, bSort, bLookup, wsFlags, nRow, cost |
| 2471 | )); |
| 2472 | |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2473 | /* If this index is the best we have seen so far, then record this |
| 2474 | ** index and its cost in the pCost structure. |
| 2475 | */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2476 | if( (!pIdx || wsFlags) && cost<pCost->rCost ){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2477 | pCost->rCost = cost; |
| 2478 | pCost->nRow = nRow; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2479 | pCost->used = used; |
| 2480 | pCost->plan.wsFlags = (wsFlags&wsFlagMask); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2481 | pCost->plan.nEq = nEq; |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2482 | pCost->plan.u.pIdx = pIdx; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2483 | } |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2484 | |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2485 | /* If there was an INDEXED BY clause, then only that one index is |
| 2486 | ** considered. */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2487 | if( pSrc->pIndex ) break; |
drh | cdaca55 | 2009-08-20 13:45:07 +0000 | [diff] [blame] | 2488 | |
| 2489 | /* Reset masks for the next index in the loop */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2490 | wsFlagMask = ~(WHERE_ROWID_EQ|WHERE_ROWID_RANGE); |
| 2491 | eqTermMask = idxEqTermMask; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2492 | } |
| 2493 | |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 2494 | /* If there is no ORDER BY clause and the SQLITE_ReverseOrder flag |
| 2495 | ** is set, then reverse the order that the index will be scanned |
| 2496 | ** in. This is used for application testing, to help find cases |
| 2497 | ** where application behaviour depends on the (undefined) order that |
| 2498 | ** SQLite outputs rows in in the absence of an ORDER BY clause. */ |
| 2499 | if( !pOrderBy && pParse->db->flags & SQLITE_ReverseOrder ){ |
| 2500 | pCost->plan.wsFlags |= WHERE_REVERSE; |
| 2501 | } |
| 2502 | |
| 2503 | assert( pOrderBy || (pCost->plan.wsFlags&WHERE_ORDERBY)==0 ); |
| 2504 | assert( pCost->plan.u.pIdx==0 || (pCost->plan.wsFlags&WHERE_ROWID_EQ)==0 ); |
| 2505 | assert( pSrc->pIndex==0 |
| 2506 | || pCost->plan.u.pIdx==0 |
| 2507 | || pCost->plan.u.pIdx==pSrc->pIndex |
| 2508 | ); |
| 2509 | |
| 2510 | WHERETRACE(("best index is: %s\n", |
| 2511 | (pCost->plan.u.pIdx ? pCost->plan.u.pIdx->zName : "ipk") |
| 2512 | )); |
| 2513 | |
| 2514 | bestOrClauseIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2515 | pCost->plan.wsFlags |= eqTermMask; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 2516 | } |
| 2517 | |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2518 | /* |
| 2519 | ** Find the query plan for accessing table pSrc->pTab. Write the |
| 2520 | ** best query plan and its cost into the WhereCost object supplied |
| 2521 | ** as the last parameter. This function may calculate the cost of |
| 2522 | ** both real and virtual table scans. |
| 2523 | */ |
| 2524 | static void bestIndex( |
| 2525 | Parse *pParse, /* The parsing context */ |
| 2526 | WhereClause *pWC, /* The WHERE clause */ |
| 2527 | struct SrcList_item *pSrc, /* The FROM clause term to search */ |
| 2528 | Bitmask notReady, /* Mask of cursors that are not available */ |
| 2529 | ExprList *pOrderBy, /* The ORDER BY clause */ |
| 2530 | WhereCost *pCost /* Lowest cost query plan */ |
| 2531 | ){ |
shane | e26fa4c | 2009-06-16 14:15:22 +0000 | [diff] [blame] | 2532 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2533 | if( IsVirtual(pSrc->pTab) ){ |
| 2534 | sqlite3_index_info *p = 0; |
| 2535 | bestVirtualIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost, &p); |
| 2536 | if( p->needToFreeIdxStr ){ |
| 2537 | sqlite3_free(p->idxStr); |
| 2538 | } |
| 2539 | sqlite3DbFree(pParse->db, p); |
shane | e26fa4c | 2009-06-16 14:15:22 +0000 | [diff] [blame] | 2540 | }else |
| 2541 | #endif |
| 2542 | { |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2543 | bestBtreeIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost); |
| 2544 | } |
| 2545 | } |
drh | b6c2989 | 2004-11-22 19:12:19 +0000 | [diff] [blame] | 2546 | |
| 2547 | /* |
drh | 2ffb118 | 2004-07-19 19:14:01 +0000 | [diff] [blame] | 2548 | ** Disable a term in the WHERE clause. Except, do not disable the term |
| 2549 | ** if it controls a LEFT OUTER JOIN and it did not originate in the ON |
| 2550 | ** or USING clause of that join. |
| 2551 | ** |
| 2552 | ** Consider the term t2.z='ok' in the following queries: |
| 2553 | ** |
| 2554 | ** (1) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x WHERE t2.z='ok' |
| 2555 | ** (2) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x AND t2.z='ok' |
| 2556 | ** (3) SELECT * FROM t1, t2 WHERE t1.a=t2.x AND t2.z='ok' |
| 2557 | ** |
drh | 23bf66d | 2004-12-14 03:34:34 +0000 | [diff] [blame] | 2558 | ** The t2.z='ok' is disabled in the in (2) because it originates |
drh | 2ffb118 | 2004-07-19 19:14:01 +0000 | [diff] [blame] | 2559 | ** in the ON clause. The term is disabled in (3) because it is not part |
| 2560 | ** of a LEFT OUTER JOIN. In (1), the term is not disabled. |
| 2561 | ** |
| 2562 | ** Disabling a term causes that term to not be tested in the inner loop |
drh | b6fb62d | 2005-09-20 08:47:20 +0000 | [diff] [blame] | 2563 | ** of the join. Disabling is an optimization. When terms are satisfied |
| 2564 | ** by indices, we disable them to prevent redundant tests in the inner |
| 2565 | ** loop. We would get the correct results if nothing were ever disabled, |
| 2566 | ** but joins might run a little slower. The trick is to disable as much |
| 2567 | ** as we can without disabling too much. If we disabled in (1), we'd get |
| 2568 | ** the wrong answer. See ticket #813. |
drh | 2ffb118 | 2004-07-19 19:14:01 +0000 | [diff] [blame] | 2569 | */ |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 2570 | static void disableTerm(WhereLevel *pLevel, WhereTerm *pTerm){ |
| 2571 | if( pTerm |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 2572 | && ALWAYS((pTerm->wtFlags & TERM_CODED)==0) |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 2573 | && (pLevel->iLeftJoin==0 || ExprHasProperty(pTerm->pExpr, EP_FromJoin)) |
| 2574 | ){ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 2575 | pTerm->wtFlags |= TERM_CODED; |
drh | 45b1ee4 | 2005-08-02 17:48:22 +0000 | [diff] [blame] | 2576 | if( pTerm->iParent>=0 ){ |
| 2577 | WhereTerm *pOther = &pTerm->pWC->a[pTerm->iParent]; |
| 2578 | if( (--pOther->nChild)==0 ){ |
drh | ed37800 | 2005-07-28 23:12:08 +0000 | [diff] [blame] | 2579 | disableTerm(pLevel, pOther); |
| 2580 | } |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 2581 | } |
drh | 2ffb118 | 2004-07-19 19:14:01 +0000 | [diff] [blame] | 2582 | } |
| 2583 | } |
| 2584 | |
| 2585 | /* |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 2586 | ** Code an OP_Affinity opcode to apply the column affinity string zAff |
| 2587 | ** to the n registers starting at base. |
| 2588 | ** |
| 2589 | ** Buffer zAff was allocated using sqlite3DbMalloc(). It is the |
| 2590 | ** responsibility of this function to arrange for it to be eventually |
| 2591 | ** freed using sqlite3DbFree(). |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 2592 | */ |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 2593 | static void codeApplyAffinity(Parse *pParse, int base, int n, char *zAff){ |
| 2594 | Vdbe *v = pParse->pVdbe; |
| 2595 | assert( v!=0 ); |
| 2596 | sqlite3VdbeAddOp2(v, OP_Affinity, base, n); |
| 2597 | sqlite3VdbeChangeP4(v, -1, zAff, P4_DYNAMIC); |
| 2598 | sqlite3ExprCacheAffinityChange(pParse, base, n); |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 2599 | } |
| 2600 | |
drh | e8b9727 | 2005-07-19 22:22:12 +0000 | [diff] [blame] | 2601 | |
| 2602 | /* |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2603 | ** Generate code for a single equality term of the WHERE clause. An equality |
| 2604 | ** term can be either X=expr or X IN (...). pTerm is the term to be |
| 2605 | ** coded. |
| 2606 | ** |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2607 | ** The current value for the constraint is left in register iReg. |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2608 | ** |
| 2609 | ** For a constraint of the form X=expr, the expression is evaluated and its |
| 2610 | ** result is left on the stack. For constraints of the form X IN (...) |
| 2611 | ** this routine sets up a loop that will iterate over all values of X. |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 2612 | */ |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2613 | static int codeEqualityTerm( |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 2614 | Parse *pParse, /* The parsing context */ |
drh | e23399f | 2005-07-22 00:31:39 +0000 | [diff] [blame] | 2615 | WhereTerm *pTerm, /* The term of the WHERE clause to be coded */ |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2616 | WhereLevel *pLevel, /* When level of the FROM clause we are working on */ |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2617 | int iTarget /* Attempt to leave results in this register */ |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 2618 | ){ |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 2619 | Expr *pX = pTerm->pExpr; |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 2620 | Vdbe *v = pParse->pVdbe; |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2621 | int iReg; /* Register holding results */ |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2622 | |
danielk1977 | 2d60549 | 2008-10-01 08:43:03 +0000 | [diff] [blame] | 2623 | assert( iTarget>0 ); |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 2624 | if( pX->op==TK_EQ ){ |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2625 | iReg = sqlite3ExprCodeTarget(pParse, pX->pRight, iTarget); |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 2626 | }else if( pX->op==TK_ISNULL ){ |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2627 | iReg = iTarget; |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2628 | sqlite3VdbeAddOp2(v, OP_Null, 0, iReg); |
danielk1977 | b3bce66 | 2005-01-29 08:32:43 +0000 | [diff] [blame] | 2629 | #ifndef SQLITE_OMIT_SUBQUERY |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 2630 | }else{ |
danielk1977 | 9a96b66 | 2007-11-29 17:05:18 +0000 | [diff] [blame] | 2631 | int eType; |
danielk1977 | b3bce66 | 2005-01-29 08:32:43 +0000 | [diff] [blame] | 2632 | int iTab; |
drh | 72e8fa4 | 2007-03-28 14:30:06 +0000 | [diff] [blame] | 2633 | struct InLoop *pIn; |
danielk1977 | b3bce66 | 2005-01-29 08:32:43 +0000 | [diff] [blame] | 2634 | |
drh | 50b3996 | 2006-10-28 00:28:09 +0000 | [diff] [blame] | 2635 | assert( pX->op==TK_IN ); |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2636 | iReg = iTarget; |
danielk1977 | 0cdc022 | 2008-06-26 18:04:03 +0000 | [diff] [blame] | 2637 | eType = sqlite3FindInIndex(pParse, pX, 0); |
danielk1977 | b3bce66 | 2005-01-29 08:32:43 +0000 | [diff] [blame] | 2638 | iTab = pX->iTable; |
drh | 66a5167 | 2008-01-03 00:01:23 +0000 | [diff] [blame] | 2639 | sqlite3VdbeAddOp2(v, OP_Rewind, iTab, 0); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2640 | assert( pLevel->plan.wsFlags & WHERE_IN_ABLE ); |
| 2641 | if( pLevel->u.in.nIn==0 ){ |
drh | b3190c1 | 2008-12-08 21:37:14 +0000 | [diff] [blame] | 2642 | pLevel->addrNxt = sqlite3VdbeMakeLabel(v); |
drh | 72e8fa4 | 2007-03-28 14:30:06 +0000 | [diff] [blame] | 2643 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2644 | pLevel->u.in.nIn++; |
| 2645 | pLevel->u.in.aInLoop = |
| 2646 | sqlite3DbReallocOrFree(pParse->db, pLevel->u.in.aInLoop, |
| 2647 | sizeof(pLevel->u.in.aInLoop[0])*pLevel->u.in.nIn); |
| 2648 | pIn = pLevel->u.in.aInLoop; |
drh | 72e8fa4 | 2007-03-28 14:30:06 +0000 | [diff] [blame] | 2649 | if( pIn ){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2650 | pIn += pLevel->u.in.nIn - 1; |
drh | 72e8fa4 | 2007-03-28 14:30:06 +0000 | [diff] [blame] | 2651 | pIn->iCur = iTab; |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2652 | if( eType==IN_INDEX_ROWID ){ |
drh | b3190c1 | 2008-12-08 21:37:14 +0000 | [diff] [blame] | 2653 | pIn->addrInTop = sqlite3VdbeAddOp2(v, OP_Rowid, iTab, iReg); |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2654 | }else{ |
drh | b3190c1 | 2008-12-08 21:37:14 +0000 | [diff] [blame] | 2655 | pIn->addrInTop = sqlite3VdbeAddOp3(v, OP_Column, iTab, 0, iReg); |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2656 | } |
| 2657 | sqlite3VdbeAddOp1(v, OP_IsNull, iReg); |
drh | a611040 | 2005-07-28 20:51:19 +0000 | [diff] [blame] | 2658 | }else{ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2659 | pLevel->u.in.nIn = 0; |
drh | e23399f | 2005-07-22 00:31:39 +0000 | [diff] [blame] | 2660 | } |
danielk1977 | b3bce66 | 2005-01-29 08:32:43 +0000 | [diff] [blame] | 2661 | #endif |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 2662 | } |
drh | 0fcef5e | 2005-07-19 17:38:22 +0000 | [diff] [blame] | 2663 | disableTerm(pLevel, pTerm); |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2664 | return iReg; |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 2665 | } |
| 2666 | |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2667 | /* |
| 2668 | ** Generate code that will evaluate all == and IN constraints for an |
| 2669 | ** index. The values for all constraints are left on the stack. |
| 2670 | ** |
| 2671 | ** For example, consider table t1(a,b,c,d,e,f) with index i1(a,b,c). |
| 2672 | ** Suppose the WHERE clause is this: a==5 AND b IN (1,2,3) AND c>5 AND c<10 |
| 2673 | ** The index has as many as three equality constraints, but in this |
| 2674 | ** example, the third "c" value is an inequality. So only two |
| 2675 | ** constraints are coded. This routine will generate code to evaluate |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 2676 | ** a==5 and b IN (1,2,3). The current values for a and b will be stored |
| 2677 | ** in consecutive registers and the index of the first register is returned. |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2678 | ** |
| 2679 | ** In the example above nEq==2. But this subroutine works for any value |
| 2680 | ** of nEq including 0. If nEq==0, this routine is nearly a no-op. |
| 2681 | ** The only thing it does is allocate the pLevel->iMem memory cell. |
| 2682 | ** |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 2683 | ** This routine always allocates at least one memory cell and returns |
| 2684 | ** the index of that memory cell. The code that |
| 2685 | ** calls this routine will use that memory cell to store the termination |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2686 | ** key value of the loop. If one or more IN operators appear, then |
| 2687 | ** this routine allocates an additional nEq memory cells for internal |
| 2688 | ** use. |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 2689 | ** |
| 2690 | ** Before returning, *pzAff is set to point to a buffer containing a |
| 2691 | ** copy of the column affinity string of the index allocated using |
| 2692 | ** sqlite3DbMalloc(). Except, entries in the copy of the string associated |
| 2693 | ** with equality constraints that use NONE affinity are set to |
| 2694 | ** SQLITE_AFF_NONE. This is to deal with SQL such as the following: |
| 2695 | ** |
| 2696 | ** CREATE TABLE t1(a TEXT PRIMARY KEY, b); |
| 2697 | ** SELECT ... FROM t1 AS t2, t1 WHERE t1.a = t2.b; |
| 2698 | ** |
| 2699 | ** In the example above, the index on t1(a) has TEXT affinity. But since |
| 2700 | ** the right hand side of the equality constraint (t2.b) has NONE affinity, |
| 2701 | ** no conversion should be attempted before using a t2.b value as part of |
| 2702 | ** a key to search the index. Hence the first byte in the returned affinity |
| 2703 | ** string in this example would be set to SQLITE_AFF_NONE. |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2704 | */ |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2705 | static int codeAllEqualityTerms( |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2706 | Parse *pParse, /* Parsing context */ |
| 2707 | WhereLevel *pLevel, /* Which nested loop of the FROM we are coding */ |
| 2708 | WhereClause *pWC, /* The WHERE clause */ |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2709 | Bitmask notReady, /* Which parts of FROM have not yet been coded */ |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 2710 | int nExtraReg, /* Number of extra registers to allocate */ |
| 2711 | char **pzAff /* OUT: Set to point to affinity string */ |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2712 | ){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2713 | int nEq = pLevel->plan.nEq; /* The number of == or IN constraints to code */ |
| 2714 | Vdbe *v = pParse->pVdbe; /* The vm under construction */ |
| 2715 | Index *pIdx; /* The index being used for this loop */ |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2716 | int iCur = pLevel->iTabCur; /* The cursor of the table */ |
| 2717 | WhereTerm *pTerm; /* A single constraint term */ |
| 2718 | int j; /* Loop counter */ |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2719 | int regBase; /* Base register */ |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 2720 | int nReg; /* Number of registers to allocate */ |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 2721 | char *zAff; /* Affinity string to return */ |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2722 | |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2723 | /* This module is only called on query plans that use an index. */ |
| 2724 | assert( pLevel->plan.wsFlags & WHERE_INDEXED ); |
| 2725 | pIdx = pLevel->plan.u.pIdx; |
| 2726 | |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2727 | /* Figure out how many memory cells we will need then allocate them. |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2728 | */ |
drh | 700a226 | 2008-12-17 19:22:15 +0000 | [diff] [blame] | 2729 | regBase = pParse->nMem + 1; |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 2730 | nReg = pLevel->plan.nEq + nExtraReg; |
| 2731 | pParse->nMem += nReg; |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2732 | |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 2733 | zAff = sqlite3DbStrDup(pParse->db, sqlite3IndexAffinityStr(v, pIdx)); |
| 2734 | if( !zAff ){ |
| 2735 | pParse->db->mallocFailed = 1; |
| 2736 | } |
| 2737 | |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2738 | /* Evaluate the equality constraints |
| 2739 | */ |
drh | c49de5d | 2007-01-19 01:06:01 +0000 | [diff] [blame] | 2740 | assert( pIdx->nColumn>=nEq ); |
| 2741 | for(j=0; j<nEq; j++){ |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2742 | int r1; |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2743 | int k = pIdx->aiColumn[j]; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2744 | pTerm = findTerm(pWC, iCur, k, notReady, pLevel->plan.wsFlags, pIdx); |
drh | 34004ce | 2008-07-11 16:15:17 +0000 | [diff] [blame] | 2745 | if( NEVER(pTerm==0) ) break; |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 2746 | assert( (pTerm->wtFlags & TERM_CODED)==0 ); |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2747 | r1 = codeEqualityTerm(pParse, pTerm, pLevel, regBase+j); |
| 2748 | if( r1!=regBase+j ){ |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 2749 | if( nReg==1 ){ |
| 2750 | sqlite3ReleaseTempReg(pParse, regBase); |
| 2751 | regBase = r1; |
| 2752 | }else{ |
| 2753 | sqlite3VdbeAddOp2(v, OP_SCopy, r1, regBase+j); |
| 2754 | } |
drh | 678ccce | 2008-03-31 18:19:54 +0000 | [diff] [blame] | 2755 | } |
drh | 981642f | 2008-04-19 14:40:43 +0000 | [diff] [blame] | 2756 | testcase( pTerm->eOperator & WO_ISNULL ); |
| 2757 | testcase( pTerm->eOperator & WO_IN ); |
drh | 72e8fa4 | 2007-03-28 14:30:06 +0000 | [diff] [blame] | 2758 | if( (pTerm->eOperator & (WO_ISNULL|WO_IN))==0 ){ |
drh | b3190c1 | 2008-12-08 21:37:14 +0000 | [diff] [blame] | 2759 | sqlite3VdbeAddOp2(v, OP_IsNull, regBase+j, pLevel->addrBrk); |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 2760 | if( zAff |
| 2761 | && sqlite3CompareAffinity(pTerm->pExpr->pRight, zAff[j])==SQLITE_AFF_NONE |
| 2762 | ){ |
| 2763 | zAff[j] = SQLITE_AFF_NONE; |
| 2764 | } |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2765 | } |
| 2766 | } |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 2767 | *pzAff = zAff; |
drh | 1db639c | 2008-01-17 02:36:28 +0000 | [diff] [blame] | 2768 | return regBase; |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 2769 | } |
| 2770 | |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2771 | /* |
| 2772 | ** Generate code for the start of the iLevel-th loop in the WHERE clause |
| 2773 | ** implementation described by pWInfo. |
| 2774 | */ |
| 2775 | static Bitmask codeOneLoopStart( |
| 2776 | WhereInfo *pWInfo, /* Complete information about the WHERE clause */ |
| 2777 | int iLevel, /* Which level of pWInfo->a[] should be coded */ |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 2778 | u16 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2779 | Bitmask notReady /* Which tables are currently available */ |
| 2780 | ){ |
| 2781 | int j, k; /* Loop counters */ |
| 2782 | int iCur; /* The VDBE cursor for the table */ |
| 2783 | int addrNxt; /* Where to jump to continue with the next IN case */ |
| 2784 | int omitTable; /* True if we use the index only */ |
| 2785 | int bRev; /* True if we need to scan in reverse order */ |
| 2786 | WhereLevel *pLevel; /* The where level to be coded */ |
| 2787 | WhereClause *pWC; /* Decomposition of the entire WHERE clause */ |
| 2788 | WhereTerm *pTerm; /* A WHERE clause term */ |
| 2789 | Parse *pParse; /* Parsing context */ |
| 2790 | Vdbe *v; /* The prepared stmt under constructions */ |
| 2791 | struct SrcList_item *pTabItem; /* FROM clause term being coded */ |
drh | 23d04d5 | 2008-12-23 23:56:22 +0000 | [diff] [blame] | 2792 | int addrBrk; /* Jump here to break out of the loop */ |
| 2793 | int addrCont; /* Jump here to continue with next cycle */ |
drh | 6149526 | 2009-04-22 15:32:59 +0000 | [diff] [blame] | 2794 | int iRowidReg = 0; /* Rowid is stored in this register, if not zero */ |
| 2795 | int iReleaseReg = 0; /* Temp register to free before returning */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2796 | |
| 2797 | pParse = pWInfo->pParse; |
| 2798 | v = pParse->pVdbe; |
| 2799 | pWC = pWInfo->pWC; |
| 2800 | pLevel = &pWInfo->a[iLevel]; |
| 2801 | pTabItem = &pWInfo->pTabList->a[pLevel->iFrom]; |
| 2802 | iCur = pTabItem->iCursor; |
| 2803 | bRev = (pLevel->plan.wsFlags & WHERE_REVERSE)!=0; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2804 | omitTable = (pLevel->plan.wsFlags & WHERE_IDX_ONLY)!=0 |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 2805 | && (wctrlFlags & WHERE_FORCE_TABLE)==0; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2806 | |
| 2807 | /* Create labels for the "break" and "continue" instructions |
| 2808 | ** for the current loop. Jump to addrBrk to break out of a loop. |
| 2809 | ** Jump to cont to go immediately to the next iteration of the |
| 2810 | ** loop. |
| 2811 | ** |
| 2812 | ** When there is an IN operator, we also have a "addrNxt" label that |
| 2813 | ** means to continue with the next IN value combination. When |
| 2814 | ** there are no IN operators in the constraints, the "addrNxt" label |
| 2815 | ** is the same as "addrBrk". |
| 2816 | */ |
| 2817 | addrBrk = pLevel->addrBrk = pLevel->addrNxt = sqlite3VdbeMakeLabel(v); |
| 2818 | addrCont = pLevel->addrCont = sqlite3VdbeMakeLabel(v); |
| 2819 | |
| 2820 | /* If this is the right table of a LEFT OUTER JOIN, allocate and |
| 2821 | ** initialize a memory cell that records if this table matches any |
| 2822 | ** row of the left table of the join. |
| 2823 | */ |
| 2824 | if( pLevel->iFrom>0 && (pTabItem[0].jointype & JT_LEFT)!=0 ){ |
| 2825 | pLevel->iLeftJoin = ++pParse->nMem; |
| 2826 | sqlite3VdbeAddOp2(v, OP_Integer, 0, pLevel->iLeftJoin); |
| 2827 | VdbeComment((v, "init LEFT JOIN no-match flag")); |
| 2828 | } |
| 2829 | |
| 2830 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
| 2831 | if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){ |
| 2832 | /* Case 0: The table is a virtual-table. Use the VFilter and VNext |
| 2833 | ** to access the data. |
| 2834 | */ |
| 2835 | int iReg; /* P3 Value for OP_VFilter */ |
| 2836 | sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx; |
| 2837 | int nConstraint = pVtabIdx->nConstraint; |
| 2838 | struct sqlite3_index_constraint_usage *aUsage = |
| 2839 | pVtabIdx->aConstraintUsage; |
| 2840 | const struct sqlite3_index_constraint *aConstraint = |
| 2841 | pVtabIdx->aConstraint; |
| 2842 | |
| 2843 | iReg = sqlite3GetTempRange(pParse, nConstraint+2); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2844 | for(j=1; j<=nConstraint; j++){ |
| 2845 | for(k=0; k<nConstraint; k++){ |
| 2846 | if( aUsage[k].argvIndex==j ){ |
| 2847 | int iTerm = aConstraint[k].iTermOffset; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2848 | sqlite3ExprCode(pParse, pWC->a[iTerm].pExpr->pRight, iReg+j+1); |
| 2849 | break; |
| 2850 | } |
| 2851 | } |
| 2852 | if( k==nConstraint ) break; |
| 2853 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2854 | sqlite3VdbeAddOp2(v, OP_Integer, pVtabIdx->idxNum, iReg); |
| 2855 | sqlite3VdbeAddOp2(v, OP_Integer, j-1, iReg+1); |
| 2856 | sqlite3VdbeAddOp4(v, OP_VFilter, iCur, addrBrk, iReg, pVtabIdx->idxStr, |
| 2857 | pVtabIdx->needToFreeIdxStr ? P4_MPRINTF : P4_STATIC); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2858 | pVtabIdx->needToFreeIdxStr = 0; |
| 2859 | for(j=0; j<nConstraint; j++){ |
| 2860 | if( aUsage[j].omit ){ |
| 2861 | int iTerm = aConstraint[j].iTermOffset; |
| 2862 | disableTerm(pLevel, &pWC->a[iTerm]); |
| 2863 | } |
| 2864 | } |
| 2865 | pLevel->op = OP_VNext; |
| 2866 | pLevel->p1 = iCur; |
| 2867 | pLevel->p2 = sqlite3VdbeCurrentAddr(v); |
drh | 23d04d5 | 2008-12-23 23:56:22 +0000 | [diff] [blame] | 2868 | sqlite3ReleaseTempRange(pParse, iReg, nConstraint+2); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2869 | }else |
| 2870 | #endif /* SQLITE_OMIT_VIRTUALTABLE */ |
| 2871 | |
| 2872 | if( pLevel->plan.wsFlags & WHERE_ROWID_EQ ){ |
| 2873 | /* Case 1: We can directly reference a single row using an |
| 2874 | ** equality comparison against the ROWID field. Or |
| 2875 | ** we reference multiple rows using a "rowid IN (...)" |
| 2876 | ** construct. |
| 2877 | */ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2878 | iReleaseReg = sqlite3GetTempReg(pParse); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2879 | pTerm = findTerm(pWC, iCur, -1, notReady, WO_EQ|WO_IN, 0); |
| 2880 | assert( pTerm!=0 ); |
| 2881 | assert( pTerm->pExpr!=0 ); |
| 2882 | assert( pTerm->leftCursor==iCur ); |
| 2883 | assert( omitTable==0 ); |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2884 | iRowidReg = codeEqualityTerm(pParse, pTerm, pLevel, iReleaseReg); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2885 | addrNxt = pLevel->addrNxt; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2886 | sqlite3VdbeAddOp2(v, OP_MustBeInt, iRowidReg, addrNxt); |
| 2887 | sqlite3VdbeAddOp3(v, OP_NotExists, iCur, addrNxt, iRowidReg); |
drh | ceea332 | 2009-04-23 13:22:42 +0000 | [diff] [blame] | 2888 | sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2889 | VdbeComment((v, "pk")); |
| 2890 | pLevel->op = OP_Noop; |
| 2891 | }else if( pLevel->plan.wsFlags & WHERE_ROWID_RANGE ){ |
| 2892 | /* Case 2: We have an inequality comparison against the ROWID field. |
| 2893 | */ |
| 2894 | int testOp = OP_Noop; |
| 2895 | int start; |
| 2896 | int memEndValue = 0; |
| 2897 | WhereTerm *pStart, *pEnd; |
| 2898 | |
| 2899 | assert( omitTable==0 ); |
| 2900 | pStart = findTerm(pWC, iCur, -1, notReady, WO_GT|WO_GE, 0); |
| 2901 | pEnd = findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE, 0); |
| 2902 | if( bRev ){ |
| 2903 | pTerm = pStart; |
| 2904 | pStart = pEnd; |
| 2905 | pEnd = pTerm; |
| 2906 | } |
| 2907 | if( pStart ){ |
| 2908 | Expr *pX; /* The expression that defines the start bound */ |
| 2909 | int r1, rTemp; /* Registers for holding the start boundary */ |
| 2910 | |
| 2911 | /* The following constant maps TK_xx codes into corresponding |
| 2912 | ** seek opcodes. It depends on a particular ordering of TK_xx |
| 2913 | */ |
| 2914 | const u8 aMoveOp[] = { |
| 2915 | /* TK_GT */ OP_SeekGt, |
| 2916 | /* TK_LE */ OP_SeekLe, |
| 2917 | /* TK_LT */ OP_SeekLt, |
| 2918 | /* TK_GE */ OP_SeekGe |
| 2919 | }; |
| 2920 | assert( TK_LE==TK_GT+1 ); /* Make sure the ordering.. */ |
| 2921 | assert( TK_LT==TK_GT+2 ); /* ... of the TK_xx values... */ |
| 2922 | assert( TK_GE==TK_GT+3 ); /* ... is correcct. */ |
| 2923 | |
| 2924 | pX = pStart->pExpr; |
| 2925 | assert( pX!=0 ); |
| 2926 | assert( pStart->leftCursor==iCur ); |
| 2927 | r1 = sqlite3ExprCodeTemp(pParse, pX->pRight, &rTemp); |
| 2928 | sqlite3VdbeAddOp3(v, aMoveOp[pX->op-TK_GT], iCur, addrBrk, r1); |
| 2929 | VdbeComment((v, "pk")); |
| 2930 | sqlite3ExprCacheAffinityChange(pParse, r1, 1); |
| 2931 | sqlite3ReleaseTempReg(pParse, rTemp); |
| 2932 | disableTerm(pLevel, pStart); |
| 2933 | }else{ |
| 2934 | sqlite3VdbeAddOp2(v, bRev ? OP_Last : OP_Rewind, iCur, addrBrk); |
| 2935 | } |
| 2936 | if( pEnd ){ |
| 2937 | Expr *pX; |
| 2938 | pX = pEnd->pExpr; |
| 2939 | assert( pX!=0 ); |
| 2940 | assert( pEnd->leftCursor==iCur ); |
| 2941 | memEndValue = ++pParse->nMem; |
| 2942 | sqlite3ExprCode(pParse, pX->pRight, memEndValue); |
| 2943 | if( pX->op==TK_LT || pX->op==TK_GT ){ |
| 2944 | testOp = bRev ? OP_Le : OP_Ge; |
| 2945 | }else{ |
| 2946 | testOp = bRev ? OP_Lt : OP_Gt; |
| 2947 | } |
| 2948 | disableTerm(pLevel, pEnd); |
| 2949 | } |
| 2950 | start = sqlite3VdbeCurrentAddr(v); |
| 2951 | pLevel->op = bRev ? OP_Prev : OP_Next; |
| 2952 | pLevel->p1 = iCur; |
| 2953 | pLevel->p2 = start; |
drh | ca8c466 | 2008-12-28 20:47:02 +0000 | [diff] [blame] | 2954 | pLevel->p5 = (pStart==0 && pEnd==0) ?1:0; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2955 | if( testOp!=OP_Noop ){ |
| 2956 | iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse); |
| 2957 | sqlite3VdbeAddOp2(v, OP_Rowid, iCur, iRowidReg); |
drh | ceea332 | 2009-04-23 13:22:42 +0000 | [diff] [blame] | 2958 | sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 2959 | sqlite3VdbeAddOp3(v, testOp, memEndValue, addrBrk, iRowidReg); |
| 2960 | sqlite3VdbeChangeP5(v, SQLITE_AFF_NUMERIC | SQLITE_JUMPIFNULL); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 2961 | } |
| 2962 | }else if( pLevel->plan.wsFlags & (WHERE_COLUMN_RANGE|WHERE_COLUMN_EQ) ){ |
| 2963 | /* Case 3: A scan using an index. |
| 2964 | ** |
| 2965 | ** The WHERE clause may contain zero or more equality |
| 2966 | ** terms ("==" or "IN" operators) that refer to the N |
| 2967 | ** left-most columns of the index. It may also contain |
| 2968 | ** inequality constraints (>, <, >= or <=) on the indexed |
| 2969 | ** column that immediately follows the N equalities. Only |
| 2970 | ** the right-most column can be an inequality - the rest must |
| 2971 | ** use the "==" and "IN" operators. For example, if the |
| 2972 | ** index is on (x,y,z), then the following clauses are all |
| 2973 | ** optimized: |
| 2974 | ** |
| 2975 | ** x=5 |
| 2976 | ** x=5 AND y=10 |
| 2977 | ** x=5 AND y<10 |
| 2978 | ** x=5 AND y>5 AND y<10 |
| 2979 | ** x=5 AND y=5 AND z<=10 |
| 2980 | ** |
| 2981 | ** The z<10 term of the following cannot be used, only |
| 2982 | ** the x=5 term: |
| 2983 | ** |
| 2984 | ** x=5 AND z<10 |
| 2985 | ** |
| 2986 | ** N may be zero if there are inequality constraints. |
| 2987 | ** If there are no inequality constraints, then N is at |
| 2988 | ** least one. |
| 2989 | ** |
| 2990 | ** This case is also used when there are no WHERE clause |
| 2991 | ** constraints but an index is selected anyway, in order |
| 2992 | ** to force the output order to conform to an ORDER BY. |
| 2993 | */ |
| 2994 | int aStartOp[] = { |
| 2995 | 0, |
| 2996 | 0, |
| 2997 | OP_Rewind, /* 2: (!start_constraints && startEq && !bRev) */ |
| 2998 | OP_Last, /* 3: (!start_constraints && startEq && bRev) */ |
| 2999 | OP_SeekGt, /* 4: (start_constraints && !startEq && !bRev) */ |
| 3000 | OP_SeekLt, /* 5: (start_constraints && !startEq && bRev) */ |
| 3001 | OP_SeekGe, /* 6: (start_constraints && startEq && !bRev) */ |
| 3002 | OP_SeekLe /* 7: (start_constraints && startEq && bRev) */ |
| 3003 | }; |
| 3004 | int aEndOp[] = { |
| 3005 | OP_Noop, /* 0: (!end_constraints) */ |
| 3006 | OP_IdxGE, /* 1: (end_constraints && !bRev) */ |
| 3007 | OP_IdxLT /* 2: (end_constraints && bRev) */ |
| 3008 | }; |
| 3009 | int nEq = pLevel->plan.nEq; |
| 3010 | int isMinQuery = 0; /* If this is an optimized SELECT min(x).. */ |
| 3011 | int regBase; /* Base register holding constraint values */ |
| 3012 | int r1; /* Temp register */ |
| 3013 | WhereTerm *pRangeStart = 0; /* Inequality constraint at range start */ |
| 3014 | WhereTerm *pRangeEnd = 0; /* Inequality constraint at range end */ |
| 3015 | int startEq; /* True if range start uses ==, >= or <= */ |
| 3016 | int endEq; /* True if range end uses ==, >= or <= */ |
| 3017 | int start_constraints; /* Start of range is constrained */ |
| 3018 | int nConstraint; /* Number of constraint terms */ |
| 3019 | Index *pIdx; /* The index we will be using */ |
| 3020 | int iIdxCur; /* The VDBE cursor for the index */ |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3021 | int nExtraReg = 0; /* Number of extra registers needed */ |
| 3022 | int op; /* Instruction opcode */ |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 3023 | char *zAff; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3024 | |
| 3025 | pIdx = pLevel->plan.u.pIdx; |
| 3026 | iIdxCur = pLevel->iIdxCur; |
| 3027 | k = pIdx->aiColumn[nEq]; /* Column for inequality constraints */ |
| 3028 | |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3029 | /* If this loop satisfies a sort order (pOrderBy) request that |
| 3030 | ** was passed to this function to implement a "SELECT min(x) ..." |
| 3031 | ** query, then the caller will only allow the loop to run for |
| 3032 | ** a single iteration. This means that the first row returned |
| 3033 | ** should not have a NULL value stored in 'x'. If column 'x' is |
| 3034 | ** the first one after the nEq equality constraints in the index, |
| 3035 | ** this requires some special handling. |
| 3036 | */ |
| 3037 | if( (wctrlFlags&WHERE_ORDERBY_MIN)!=0 |
| 3038 | && (pLevel->plan.wsFlags&WHERE_ORDERBY) |
| 3039 | && (pIdx->nColumn>nEq) |
| 3040 | ){ |
| 3041 | /* assert( pOrderBy->nExpr==1 ); */ |
| 3042 | /* assert( pOrderBy->a[0].pExpr->iColumn==pIdx->aiColumn[nEq] ); */ |
| 3043 | isMinQuery = 1; |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3044 | nExtraReg = 1; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3045 | } |
| 3046 | |
| 3047 | /* Find any inequality constraint terms for the start and end |
| 3048 | ** of the range. |
| 3049 | */ |
| 3050 | if( pLevel->plan.wsFlags & WHERE_TOP_LIMIT ){ |
| 3051 | pRangeEnd = findTerm(pWC, iCur, k, notReady, (WO_LT|WO_LE), pIdx); |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3052 | nExtraReg = 1; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3053 | } |
| 3054 | if( pLevel->plan.wsFlags & WHERE_BTM_LIMIT ){ |
| 3055 | pRangeStart = findTerm(pWC, iCur, k, notReady, (WO_GT|WO_GE), pIdx); |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3056 | nExtraReg = 1; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3057 | } |
| 3058 | |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3059 | /* Generate code to evaluate all constraint terms using == or IN |
| 3060 | ** and store the values of those terms in an array of registers |
| 3061 | ** starting at regBase. |
| 3062 | */ |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 3063 | regBase = codeAllEqualityTerms( |
| 3064 | pParse, pLevel, pWC, notReady, nExtraReg, &zAff |
| 3065 | ); |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3066 | addrNxt = pLevel->addrNxt; |
| 3067 | |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3068 | /* If we are doing a reverse order scan on an ascending index, or |
| 3069 | ** a forward order scan on a descending index, interchange the |
| 3070 | ** start and end terms (pRangeStart and pRangeEnd). |
| 3071 | */ |
| 3072 | if( bRev==(pIdx->aSortOrder[nEq]==SQLITE_SO_ASC) ){ |
| 3073 | SWAP(WhereTerm *, pRangeEnd, pRangeStart); |
| 3074 | } |
| 3075 | |
| 3076 | testcase( pRangeStart && pRangeStart->eOperator & WO_LE ); |
| 3077 | testcase( pRangeStart && pRangeStart->eOperator & WO_GE ); |
| 3078 | testcase( pRangeEnd && pRangeEnd->eOperator & WO_LE ); |
| 3079 | testcase( pRangeEnd && pRangeEnd->eOperator & WO_GE ); |
| 3080 | startEq = !pRangeStart || pRangeStart->eOperator & (WO_LE|WO_GE); |
| 3081 | endEq = !pRangeEnd || pRangeEnd->eOperator & (WO_LE|WO_GE); |
| 3082 | start_constraints = pRangeStart || nEq>0; |
| 3083 | |
| 3084 | /* Seek the index cursor to the start of the range. */ |
| 3085 | nConstraint = nEq; |
| 3086 | if( pRangeStart ){ |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 3087 | Expr *pRight = pRangeStart->pExpr->pRight; |
| 3088 | sqlite3ExprCode(pParse, pRight, regBase+nEq); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3089 | sqlite3VdbeAddOp2(v, OP_IsNull, regBase+nEq, addrNxt); |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 3090 | if( zAff |
| 3091 | && sqlite3CompareAffinity(pRight, zAff[nConstraint])==SQLITE_AFF_NONE |
| 3092 | ){ |
| 3093 | /* Since the comparison is to be performed with no conversions applied |
| 3094 | ** to the operands, set the affinity to apply to pRight to |
| 3095 | ** SQLITE_AFF_NONE. */ |
| 3096 | zAff[nConstraint] = SQLITE_AFF_NONE; |
| 3097 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3098 | nConstraint++; |
| 3099 | }else if( isMinQuery ){ |
| 3100 | sqlite3VdbeAddOp2(v, OP_Null, 0, regBase+nEq); |
| 3101 | nConstraint++; |
| 3102 | startEq = 0; |
| 3103 | start_constraints = 1; |
| 3104 | } |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 3105 | codeApplyAffinity(pParse, regBase, nConstraint, zAff); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3106 | op = aStartOp[(start_constraints<<2) + (startEq<<1) + bRev]; |
| 3107 | assert( op!=0 ); |
| 3108 | testcase( op==OP_Rewind ); |
| 3109 | testcase( op==OP_Last ); |
| 3110 | testcase( op==OP_SeekGt ); |
| 3111 | testcase( op==OP_SeekGe ); |
| 3112 | testcase( op==OP_SeekLe ); |
| 3113 | testcase( op==OP_SeekLt ); |
| 3114 | sqlite3VdbeAddOp4(v, op, iIdxCur, addrNxt, regBase, |
| 3115 | SQLITE_INT_TO_PTR(nConstraint), P4_INT32); |
| 3116 | |
| 3117 | /* Load the value for the inequality constraint at the end of the |
| 3118 | ** range (if any). |
| 3119 | */ |
| 3120 | nConstraint = nEq; |
| 3121 | if( pRangeEnd ){ |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 3122 | Expr *pRight = pRangeEnd->pExpr->pRight; |
drh | ceea332 | 2009-04-23 13:22:42 +0000 | [diff] [blame] | 3123 | sqlite3ExprCacheRemove(pParse, regBase+nEq); |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 3124 | sqlite3ExprCode(pParse, pRight, regBase+nEq); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3125 | sqlite3VdbeAddOp2(v, OP_IsNull, regBase+nEq, addrNxt); |
dan | 69f8bb9 | 2009-08-13 19:21:16 +0000 | [diff] [blame] | 3126 | zAff = sqlite3DbStrDup(pParse->db, zAff); |
| 3127 | if( zAff |
| 3128 | && sqlite3CompareAffinity(pRight, zAff[nConstraint])==SQLITE_AFF_NONE |
| 3129 | ){ |
| 3130 | /* Since the comparison is to be performed with no conversions applied |
| 3131 | ** to the operands, set the affinity to apply to pRight to |
| 3132 | ** SQLITE_AFF_NONE. */ |
| 3133 | zAff[nConstraint] = SQLITE_AFF_NONE; |
| 3134 | } |
| 3135 | codeApplyAffinity(pParse, regBase, nEq+1, zAff); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3136 | nConstraint++; |
| 3137 | } |
| 3138 | |
| 3139 | /* Top of the loop body */ |
| 3140 | pLevel->p2 = sqlite3VdbeCurrentAddr(v); |
| 3141 | |
| 3142 | /* Check if the index cursor is past the end of the range. */ |
| 3143 | op = aEndOp[(pRangeEnd || nEq) * (1 + bRev)]; |
| 3144 | testcase( op==OP_Noop ); |
| 3145 | testcase( op==OP_IdxGE ); |
| 3146 | testcase( op==OP_IdxLT ); |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3147 | if( op!=OP_Noop ){ |
| 3148 | sqlite3VdbeAddOp4(v, op, iIdxCur, addrNxt, regBase, |
| 3149 | SQLITE_INT_TO_PTR(nConstraint), P4_INT32); |
| 3150 | sqlite3VdbeChangeP5(v, endEq!=bRev ?1:0); |
| 3151 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3152 | |
| 3153 | /* If there are inequality constraints, check that the value |
| 3154 | ** of the table column that the inequality contrains is not NULL. |
| 3155 | ** If it is, jump to the next iteration of the loop. |
| 3156 | */ |
| 3157 | r1 = sqlite3GetTempReg(pParse); |
| 3158 | testcase( pLevel->plan.wsFlags & WHERE_BTM_LIMIT ); |
| 3159 | testcase( pLevel->plan.wsFlags & WHERE_TOP_LIMIT ); |
| 3160 | if( pLevel->plan.wsFlags & (WHERE_BTM_LIMIT|WHERE_TOP_LIMIT) ){ |
| 3161 | sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, nEq, r1); |
| 3162 | sqlite3VdbeAddOp2(v, OP_IsNull, r1, addrCont); |
| 3163 | } |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3164 | sqlite3ReleaseTempReg(pParse, r1); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3165 | |
| 3166 | /* Seek the table cursor, if required */ |
drh | 23d04d5 | 2008-12-23 23:56:22 +0000 | [diff] [blame] | 3167 | disableTerm(pLevel, pRangeStart); |
| 3168 | disableTerm(pLevel, pRangeEnd); |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3169 | if( !omitTable ){ |
| 3170 | iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse); |
| 3171 | sqlite3VdbeAddOp2(v, OP_IdxRowid, iIdxCur, iRowidReg); |
drh | ceea332 | 2009-04-23 13:22:42 +0000 | [diff] [blame] | 3172 | sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3173 | sqlite3VdbeAddOp2(v, OP_Seek, iCur, iRowidReg); /* Deferred seek */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3174 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3175 | |
| 3176 | /* Record the instruction used to terminate the loop. Disable |
| 3177 | ** WHERE clause terms made redundant by the index range scan. |
| 3178 | */ |
| 3179 | pLevel->op = bRev ? OP_Prev : OP_Next; |
| 3180 | pLevel->p1 = iIdxCur; |
drh | dd5f5a6 | 2008-12-23 13:35:23 +0000 | [diff] [blame] | 3181 | }else |
| 3182 | |
drh | 23d04d5 | 2008-12-23 23:56:22 +0000 | [diff] [blame] | 3183 | #ifndef SQLITE_OMIT_OR_OPTIMIZATION |
drh | dd5f5a6 | 2008-12-23 13:35:23 +0000 | [diff] [blame] | 3184 | if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3185 | /* Case 4: Two or more separately indexed terms connected by OR |
| 3186 | ** |
| 3187 | ** Example: |
| 3188 | ** |
| 3189 | ** CREATE TABLE t1(a,b,c,d); |
| 3190 | ** CREATE INDEX i1 ON t1(a); |
| 3191 | ** CREATE INDEX i2 ON t1(b); |
| 3192 | ** CREATE INDEX i3 ON t1(c); |
| 3193 | ** |
| 3194 | ** SELECT * FROM t1 WHERE a=5 OR b=7 OR (c=11 AND d=13) |
| 3195 | ** |
| 3196 | ** In the example, there are three indexed terms connected by OR. |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3197 | ** The top of the loop looks like this: |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3198 | ** |
drh | 1b26c7c | 2009-04-22 02:15:47 +0000 | [diff] [blame] | 3199 | ** Null 1 # Zero the rowset in reg 1 |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3200 | ** |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3201 | ** Then, for each indexed term, the following. The arguments to |
drh | 1b26c7c | 2009-04-22 02:15:47 +0000 | [diff] [blame] | 3202 | ** RowSetTest are such that the rowid of the current row is inserted |
| 3203 | ** into the RowSet. If it is already present, control skips the |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3204 | ** Gosub opcode and jumps straight to the code generated by WhereEnd(). |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3205 | ** |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3206 | ** sqlite3WhereBegin(<term>) |
drh | 1b26c7c | 2009-04-22 02:15:47 +0000 | [diff] [blame] | 3207 | ** RowSetTest # Insert rowid into rowset |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3208 | ** Gosub 2 A |
| 3209 | ** sqlite3WhereEnd() |
| 3210 | ** |
| 3211 | ** Following the above, code to terminate the loop. Label A, the target |
| 3212 | ** of the Gosub above, jumps to the instruction right after the Goto. |
| 3213 | ** |
drh | 1b26c7c | 2009-04-22 02:15:47 +0000 | [diff] [blame] | 3214 | ** Null 1 # Zero the rowset in reg 1 |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3215 | ** Goto B # The loop is finished. |
| 3216 | ** |
| 3217 | ** A: <loop body> # Return data, whatever. |
| 3218 | ** |
| 3219 | ** Return 2 # Jump back to the Gosub |
| 3220 | ** |
| 3221 | ** B: <after the loop> |
| 3222 | ** |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3223 | */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3224 | WhereClause *pOrWc; /* The OR-clause broken out into subterms */ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3225 | WhereTerm *pFinal; /* Final subterm within the OR-clause. */ |
drh | dd5f5a6 | 2008-12-23 13:35:23 +0000 | [diff] [blame] | 3226 | SrcList oneTab; /* Shortened table list */ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3227 | |
| 3228 | int regReturn = ++pParse->nMem; /* Register used with OP_Gosub */ |
shane | 8509570 | 2009-06-15 16:27:08 +0000 | [diff] [blame] | 3229 | int regRowset = 0; /* Register for RowSet object */ |
| 3230 | int regRowid = 0; /* Register holding rowid */ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3231 | int iLoopBody = sqlite3VdbeMakeLabel(v); /* Start of loop body */ |
| 3232 | int iRetInit; /* Address of regReturn init */ |
| 3233 | int ii; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3234 | |
| 3235 | pTerm = pLevel->plan.u.pTerm; |
| 3236 | assert( pTerm!=0 ); |
| 3237 | assert( pTerm->eOperator==WO_OR ); |
| 3238 | assert( (pTerm->wtFlags & TERM_ORINFO)!=0 ); |
| 3239 | pOrWc = &pTerm->u.pOrInfo->wc; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3240 | pFinal = &pOrWc->a[pOrWc->nTerm-1]; |
drh | 23d04d5 | 2008-12-23 23:56:22 +0000 | [diff] [blame] | 3241 | |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3242 | /* Set up a SrcList containing just the table being scanned by this loop. */ |
drh | dd5f5a6 | 2008-12-23 13:35:23 +0000 | [diff] [blame] | 3243 | oneTab.nSrc = 1; |
| 3244 | oneTab.nAlloc = 1; |
| 3245 | oneTab.a[0] = *pTabItem; |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3246 | |
drh | 1b26c7c | 2009-04-22 02:15:47 +0000 | [diff] [blame] | 3247 | /* Initialize the rowset register to contain NULL. An SQL NULL is |
| 3248 | ** equivalent to an empty rowset. |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3249 | ** |
| 3250 | ** Also initialize regReturn to contain the address of the instruction |
| 3251 | ** immediately following the OP_Return at the bottom of the loop. This |
| 3252 | ** is required in a few obscure LEFT JOIN cases where control jumps |
| 3253 | ** over the top of the loop into the body of it. In this case the |
| 3254 | ** correct response for the end-of-loop code (the OP_Return) is to |
| 3255 | ** fall through to the next instruction, just as an OP_Next does if |
| 3256 | ** called on an uninitialized cursor. |
| 3257 | */ |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 3258 | if( (wctrlFlags & WHERE_DUPLICATES_OK)==0 ){ |
| 3259 | regRowset = ++pParse->nMem; |
| 3260 | regRowid = ++pParse->nMem; |
| 3261 | sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset); |
| 3262 | } |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3263 | iRetInit = sqlite3VdbeAddOp2(v, OP_Integer, 0, regReturn); |
| 3264 | |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3265 | for(ii=0; ii<pOrWc->nTerm; ii++){ |
| 3266 | WhereTerm *pOrTerm = &pOrWc->a[ii]; |
| 3267 | if( pOrTerm->leftCursor==iCur || pOrTerm->eOperator==WO_AND ){ |
| 3268 | WhereInfo *pSubWInfo; /* Info for single OR-term scan */ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3269 | /* Loop through table entries that match term pOrTerm. */ |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 3270 | pSubWInfo = sqlite3WhereBegin(pParse, &oneTab, pOrTerm->pExpr, 0, |
| 3271 | WHERE_OMIT_OPEN | WHERE_OMIT_CLOSE | WHERE_FORCE_TABLE); |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3272 | if( pSubWInfo ){ |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 3273 | if( (wctrlFlags & WHERE_DUPLICATES_OK)==0 ){ |
| 3274 | int iSet = ((ii==pOrWc->nTerm-1)?-1:ii); |
| 3275 | int r; |
| 3276 | r = sqlite3ExprCodeGetColumn(pParse, pTabItem->pTab, -1, iCur, |
| 3277 | regRowid, 0); |
shane | 8509570 | 2009-06-15 16:27:08 +0000 | [diff] [blame] | 3278 | sqlite3VdbeAddOp4(v, OP_RowSetTest, regRowset, |
shane | 60a4b53 | 2009-05-06 18:57:09 +0000 | [diff] [blame] | 3279 | sqlite3VdbeCurrentAddr(v)+2, |
| 3280 | r, SQLITE_INT_TO_PTR(iSet), P4_INT32); |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 3281 | } |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3282 | sqlite3VdbeAddOp2(v, OP_Gosub, regReturn, iLoopBody); |
| 3283 | |
| 3284 | /* Finish the loop through table entries that match term pOrTerm. */ |
| 3285 | sqlite3WhereEnd(pSubWInfo); |
| 3286 | } |
drh | dd5f5a6 | 2008-12-23 13:35:23 +0000 | [diff] [blame] | 3287 | } |
| 3288 | } |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3289 | sqlite3VdbeChangeP1(v, iRetInit, sqlite3VdbeCurrentAddr(v)); |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 3290 | /* sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset); */ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3291 | sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrBrk); |
| 3292 | sqlite3VdbeResolveLabel(v, iLoopBody); |
| 3293 | |
| 3294 | pLevel->op = OP_Return; |
| 3295 | pLevel->p1 = regReturn; |
drh | 23d04d5 | 2008-12-23 23:56:22 +0000 | [diff] [blame] | 3296 | disableTerm(pLevel, pTerm); |
drh | dd5f5a6 | 2008-12-23 13:35:23 +0000 | [diff] [blame] | 3297 | }else |
drh | 23d04d5 | 2008-12-23 23:56:22 +0000 | [diff] [blame] | 3298 | #endif /* SQLITE_OMIT_OR_OPTIMIZATION */ |
drh | dd5f5a6 | 2008-12-23 13:35:23 +0000 | [diff] [blame] | 3299 | |
| 3300 | { |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3301 | /* Case 5: There is no usable index. We must do a complete |
| 3302 | ** scan of the entire table. |
| 3303 | */ |
drh | 699b3d4 | 2009-02-23 16:52:07 +0000 | [diff] [blame] | 3304 | static const u8 aStep[] = { OP_Next, OP_Prev }; |
| 3305 | static const u8 aStart[] = { OP_Rewind, OP_Last }; |
| 3306 | assert( bRev==0 || bRev==1 ); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3307 | assert( omitTable==0 ); |
drh | 699b3d4 | 2009-02-23 16:52:07 +0000 | [diff] [blame] | 3308 | pLevel->op = aStep[bRev]; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3309 | pLevel->p1 = iCur; |
drh | 699b3d4 | 2009-02-23 16:52:07 +0000 | [diff] [blame] | 3310 | pLevel->p2 = 1 + sqlite3VdbeAddOp2(v, aStart[bRev], iCur, addrBrk); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3311 | pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP; |
| 3312 | } |
| 3313 | notReady &= ~getMask(pWC->pMaskSet, iCur); |
| 3314 | |
| 3315 | /* Insert code to test every subexpression that can be completely |
| 3316 | ** computed using the current set of tables. |
| 3317 | */ |
| 3318 | k = 0; |
| 3319 | for(pTerm=pWC->a, j=pWC->nTerm; j>0; j--, pTerm++){ |
| 3320 | Expr *pE; |
| 3321 | testcase( pTerm->wtFlags & TERM_VIRTUAL ); |
| 3322 | testcase( pTerm->wtFlags & TERM_CODED ); |
| 3323 | if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue; |
| 3324 | if( (pTerm->prereqAll & notReady)!=0 ) continue; |
| 3325 | pE = pTerm->pExpr; |
| 3326 | assert( pE!=0 ); |
| 3327 | if( pLevel->iLeftJoin && !ExprHasProperty(pE, EP_FromJoin) ){ |
| 3328 | continue; |
| 3329 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3330 | sqlite3ExprIfFalse(pParse, pE, addrCont, SQLITE_JUMPIFNULL); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3331 | k = 1; |
| 3332 | pTerm->wtFlags |= TERM_CODED; |
| 3333 | } |
| 3334 | |
| 3335 | /* For a LEFT OUTER JOIN, generate code that will record the fact that |
| 3336 | ** at least one row of the right table has matched the left table. |
| 3337 | */ |
| 3338 | if( pLevel->iLeftJoin ){ |
| 3339 | pLevel->addrFirst = sqlite3VdbeCurrentAddr(v); |
| 3340 | sqlite3VdbeAddOp2(v, OP_Integer, 1, pLevel->iLeftJoin); |
| 3341 | VdbeComment((v, "record LEFT JOIN hit")); |
drh | ceea332 | 2009-04-23 13:22:42 +0000 | [diff] [blame] | 3342 | sqlite3ExprCacheClear(pParse); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3343 | for(pTerm=pWC->a, j=0; j<pWC->nTerm; j++, pTerm++){ |
| 3344 | testcase( pTerm->wtFlags & TERM_VIRTUAL ); |
| 3345 | testcase( pTerm->wtFlags & TERM_CODED ); |
| 3346 | if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue; |
| 3347 | if( (pTerm->prereqAll & notReady)!=0 ) continue; |
| 3348 | assert( pTerm->pExpr ); |
| 3349 | sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL); |
| 3350 | pTerm->wtFlags |= TERM_CODED; |
| 3351 | } |
| 3352 | } |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3353 | sqlite3ReleaseTempReg(pParse, iReleaseReg); |
drh | 23d04d5 | 2008-12-23 23:56:22 +0000 | [diff] [blame] | 3354 | |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3355 | return notReady; |
| 3356 | } |
| 3357 | |
drh | 549c8b6 | 2005-09-19 13:15:23 +0000 | [diff] [blame] | 3358 | #if defined(SQLITE_TEST) |
drh | 84bfda4 | 2005-07-15 13:05:21 +0000 | [diff] [blame] | 3359 | /* |
| 3360 | ** The following variable holds a text description of query plan generated |
| 3361 | ** by the most recent call to sqlite3WhereBegin(). Each call to WhereBegin |
| 3362 | ** overwrites the previous. This information is used for testing and |
| 3363 | ** analysis only. |
| 3364 | */ |
| 3365 | char sqlite3_query_plan[BMS*2*40]; /* Text of the join */ |
| 3366 | static int nQPlan = 0; /* Next free slow in _query_plan[] */ |
| 3367 | |
| 3368 | #endif /* SQLITE_TEST */ |
| 3369 | |
| 3370 | |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3371 | /* |
| 3372 | ** Free a WhereInfo structure |
| 3373 | */ |
drh | 10fe840 | 2008-10-11 16:47:35 +0000 | [diff] [blame] | 3374 | static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3375 | if( pWInfo ){ |
| 3376 | int i; |
| 3377 | for(i=0; i<pWInfo->nLevel; i++){ |
drh | 4be8b51 | 2006-06-13 23:51:34 +0000 | [diff] [blame] | 3378 | sqlite3_index_info *pInfo = pWInfo->a[i].pIdxInfo; |
| 3379 | if( pInfo ){ |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3380 | /* assert( pInfo->needToFreeIdxStr==0 || db->mallocFailed ); */ |
danielk1977 | 8044294 | 2008-12-24 11:25:39 +0000 | [diff] [blame] | 3381 | if( pInfo->needToFreeIdxStr ){ |
| 3382 | sqlite3_free(pInfo->idxStr); |
danielk1977 | be22965 | 2009-03-20 14:18:51 +0000 | [diff] [blame] | 3383 | } |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 3384 | sqlite3DbFree(db, pInfo); |
danielk1977 | be8a783 | 2006-06-13 15:00:54 +0000 | [diff] [blame] | 3385 | } |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3386 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3387 | whereClauseClear(pWInfo->pWC); |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 3388 | sqlite3DbFree(db, pWInfo); |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3389 | } |
| 3390 | } |
| 3391 | |
drh | 94a1121 | 2004-09-25 13:12:14 +0000 | [diff] [blame] | 3392 | |
| 3393 | /* |
drh | e318474 | 2002-06-19 14:27:05 +0000 | [diff] [blame] | 3394 | ** Generate the beginning of the loop used for WHERE clause processing. |
drh | acf3b98 | 2005-01-03 01:27:18 +0000 | [diff] [blame] | 3395 | ** The return value is a pointer to an opaque structure that contains |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3396 | ** information needed to terminate the loop. Later, the calling routine |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3397 | ** should invoke sqlite3WhereEnd() with the return value of this function |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3398 | ** in order to complete the WHERE clause processing. |
| 3399 | ** |
| 3400 | ** If an error occurs, this routine returns NULL. |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 3401 | ** |
| 3402 | ** The basic idea is to do a nested loop, one loop for each table in |
| 3403 | ** the FROM clause of a select. (INSERT and UPDATE statements are the |
| 3404 | ** same as a SELECT with only a single table in the FROM clause.) For |
| 3405 | ** example, if the SQL is this: |
| 3406 | ** |
| 3407 | ** SELECT * FROM t1, t2, t3 WHERE ...; |
| 3408 | ** |
| 3409 | ** Then the code generated is conceptually like the following: |
| 3410 | ** |
| 3411 | ** foreach row1 in t1 do \ Code generated |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3412 | ** foreach row2 in t2 do |-- by sqlite3WhereBegin() |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 3413 | ** foreach row3 in t3 do / |
| 3414 | ** ... |
| 3415 | ** end \ Code generated |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3416 | ** end |-- by sqlite3WhereEnd() |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 3417 | ** end / |
| 3418 | ** |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3419 | ** Note that the loops might not be nested in the order in which they |
| 3420 | ** appear in the FROM clause if a different order is better able to make |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 3421 | ** use of indices. Note also that when the IN operator appears in |
| 3422 | ** the WHERE clause, it might result in additional nested loops for |
| 3423 | ** scanning through all values on the right-hand side of the IN. |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3424 | ** |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 3425 | ** There are Btree cursors associated with each table. t1 uses cursor |
drh | 6a3ea0e | 2003-05-02 14:32:12 +0000 | [diff] [blame] | 3426 | ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. |
| 3427 | ** And so forth. This routine generates code to open those VDBE cursors |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3428 | ** and sqlite3WhereEnd() generates the code to close them. |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 3429 | ** |
drh | e6f85e7 | 2004-12-25 01:03:13 +0000 | [diff] [blame] | 3430 | ** The code that sqlite3WhereBegin() generates leaves the cursors named |
| 3431 | ** in pTabList pointing at their appropriate entries. The [...] code |
drh | f0863fe | 2005-06-12 21:35:51 +0000 | [diff] [blame] | 3432 | ** can use OP_Column and OP_Rowid opcodes on these cursors to extract |
drh | e6f85e7 | 2004-12-25 01:03:13 +0000 | [diff] [blame] | 3433 | ** data from the various tables of the loop. |
| 3434 | ** |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 3435 | ** If the WHERE clause is empty, the foreach loops must each scan their |
| 3436 | ** entire tables. Thus a three-way join is an O(N^3) operation. But if |
| 3437 | ** the tables have indices and there are terms in the WHERE clause that |
| 3438 | ** refer to those indices, a complete table scan can be avoided and the |
| 3439 | ** code will run much faster. Most of the work of this routine is checking |
| 3440 | ** to see if there are indices that can be used to speed up the loop. |
| 3441 | ** |
| 3442 | ** Terms of the WHERE clause are also used to limit which rows actually |
| 3443 | ** make it to the "..." in the middle of the loop. After each "foreach", |
| 3444 | ** terms of the WHERE clause that use only terms in that loop and outer |
| 3445 | ** loops are evaluated and if false a jump is made around all subsequent |
| 3446 | ** inner loops (or around the "..." if the test occurs within the inner- |
| 3447 | ** most loop) |
| 3448 | ** |
| 3449 | ** OUTER JOINS |
| 3450 | ** |
| 3451 | ** An outer join of tables t1 and t2 is conceptally coded as follows: |
| 3452 | ** |
| 3453 | ** foreach row1 in t1 do |
| 3454 | ** flag = 0 |
| 3455 | ** foreach row2 in t2 do |
| 3456 | ** start: |
| 3457 | ** ... |
| 3458 | ** flag = 1 |
| 3459 | ** end |
drh | e318474 | 2002-06-19 14:27:05 +0000 | [diff] [blame] | 3460 | ** if flag==0 then |
| 3461 | ** move the row2 cursor to a null row |
| 3462 | ** goto start |
| 3463 | ** fi |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 3464 | ** end |
| 3465 | ** |
drh | e318474 | 2002-06-19 14:27:05 +0000 | [diff] [blame] | 3466 | ** ORDER BY CLAUSE PROCESSING |
| 3467 | ** |
| 3468 | ** *ppOrderBy is a pointer to the ORDER BY clause of a SELECT statement, |
| 3469 | ** if there is one. If there is no ORDER BY clause or if this routine |
| 3470 | ** is called from an UPDATE or DELETE statement, then ppOrderBy is NULL. |
| 3471 | ** |
| 3472 | ** If an index can be used so that the natural output order of the table |
| 3473 | ** scan is correct for the ORDER BY clause, then that index is used and |
| 3474 | ** *ppOrderBy is set to NULL. This is an optimization that prevents an |
| 3475 | ** unnecessary sort of the result set if an index appropriate for the |
| 3476 | ** ORDER BY clause already exists. |
| 3477 | ** |
| 3478 | ** If the where clause loops cannot be arranged to provide the correct |
| 3479 | ** output order, then the *ppOrderBy is unchanged. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3480 | */ |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3481 | WhereInfo *sqlite3WhereBegin( |
danielk1977 | ed326d7 | 2004-11-16 15:50:19 +0000 | [diff] [blame] | 3482 | Parse *pParse, /* The parser context */ |
| 3483 | SrcList *pTabList, /* A list of all tables to be scanned */ |
| 3484 | Expr *pWhere, /* The WHERE clause */ |
danielk1977 | a9d1ccb | 2008-01-05 17:39:29 +0000 | [diff] [blame] | 3485 | ExprList **ppOrderBy, /* An ORDER BY clause, or NULL */ |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 3486 | u16 wctrlFlags /* One of the WHERE_* flags defined in sqliteInt.h */ |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3487 | ){ |
| 3488 | int i; /* Loop counter */ |
danielk1977 | be22965 | 2009-03-20 14:18:51 +0000 | [diff] [blame] | 3489 | int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3490 | WhereInfo *pWInfo; /* Will become the return value of this function */ |
| 3491 | Vdbe *v = pParse->pVdbe; /* The virtual database engine */ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 3492 | Bitmask notReady; /* Cursors that are not yet positioned */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3493 | WhereMaskSet *pMaskSet; /* The expression mask set */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3494 | WhereClause *pWC; /* Decomposition of the WHERE clause */ |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3495 | struct SrcList_item *pTabItem; /* A single entry from pTabList */ |
| 3496 | WhereLevel *pLevel; /* A single level in the pWInfo list */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3497 | int iFrom; /* First unused FROM clause element */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3498 | int andFlags; /* AND-ed combination of all pWC->a[].wtFlags */ |
drh | 1743575 | 2007-08-16 04:30:38 +0000 | [diff] [blame] | 3499 | sqlite3 *db; /* Database connection */ |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3500 | |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3501 | /* The number of tables in the FROM clause is limited by the number of |
drh | 1398ad3 | 2005-01-19 23:24:50 +0000 | [diff] [blame] | 3502 | ** bits in a Bitmask |
| 3503 | */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3504 | if( pTabList->nSrc>BMS ){ |
| 3505 | sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); |
drh | 1398ad3 | 2005-01-19 23:24:50 +0000 | [diff] [blame] | 3506 | return 0; |
| 3507 | } |
| 3508 | |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3509 | /* Allocate and initialize the WhereInfo structure that will become the |
danielk1977 | be22965 | 2009-03-20 14:18:51 +0000 | [diff] [blame] | 3510 | ** return value. A single allocation is used to store the WhereInfo |
| 3511 | ** struct, the contents of WhereInfo.a[], the WhereClause structure |
| 3512 | ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte |
| 3513 | ** field (type Bitmask) it must be aligned on an 8-byte boundary on |
| 3514 | ** some architectures. Hence the ROUND8() below. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3515 | */ |
drh | 1743575 | 2007-08-16 04:30:38 +0000 | [diff] [blame] | 3516 | db = pParse->db; |
danielk1977 | be22965 | 2009-03-20 14:18:51 +0000 | [diff] [blame] | 3517 | nByteWInfo = ROUND8(sizeof(WhereInfo)+(pTabList->nSrc-1)*sizeof(WhereLevel)); |
| 3518 | pWInfo = sqlite3DbMallocZero(db, |
| 3519 | nByteWInfo + |
| 3520 | sizeof(WhereClause) + |
| 3521 | sizeof(WhereMaskSet) |
| 3522 | ); |
drh | 1743575 | 2007-08-16 04:30:38 +0000 | [diff] [blame] | 3523 | if( db->mallocFailed ){ |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 3524 | goto whereBeginError; |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3525 | } |
danielk1977 | 70b6d57 | 2006-06-19 04:49:34 +0000 | [diff] [blame] | 3526 | pWInfo->nLevel = pTabList->nSrc; |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3527 | pWInfo->pParse = pParse; |
| 3528 | pWInfo->pTabList = pTabList; |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3529 | pWInfo->iBreak = sqlite3VdbeMakeLabel(v); |
danielk1977 | be22965 | 2009-03-20 14:18:51 +0000 | [diff] [blame] | 3530 | pWInfo->pWC = pWC = (WhereClause *)&((u8 *)pWInfo)[nByteWInfo]; |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3531 | pWInfo->wctrlFlags = wctrlFlags; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3532 | pMaskSet = (WhereMaskSet*)&pWC[1]; |
drh | 08192d5 | 2002-04-30 19:20:28 +0000 | [diff] [blame] | 3533 | |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3534 | /* Split the WHERE clause into separate subexpressions where each |
| 3535 | ** subexpression is separated by an AND operator. |
| 3536 | */ |
| 3537 | initMaskSet(pMaskSet); |
| 3538 | whereClauseInit(pWC, pParse, pMaskSet); |
| 3539 | sqlite3ExprCodeConstants(pParse, pWhere); |
| 3540 | whereSplit(pWC, pWhere, TK_AND); |
| 3541 | |
drh | 08192d5 | 2002-04-30 19:20:28 +0000 | [diff] [blame] | 3542 | /* Special case: a WHERE clause that is constant. Evaluate the |
| 3543 | ** expression and either jump over all of the code or fall thru. |
| 3544 | */ |
drh | 0a16837 | 2007-06-08 00:20:47 +0000 | [diff] [blame] | 3545 | if( pWhere && (pTabList->nSrc==0 || sqlite3ExprIsConstantNotJoin(pWhere)) ){ |
drh | 3557335 | 2008-01-08 23:54:25 +0000 | [diff] [blame] | 3546 | sqlite3ExprIfFalse(pParse, pWhere, pWInfo->iBreak, SQLITE_JUMPIFNULL); |
drh | df199a2 | 2002-06-14 22:38:41 +0000 | [diff] [blame] | 3547 | pWhere = 0; |
drh | 08192d5 | 2002-04-30 19:20:28 +0000 | [diff] [blame] | 3548 | } |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3549 | |
drh | 42165be | 2008-03-26 14:56:34 +0000 | [diff] [blame] | 3550 | /* Assign a bit from the bitmask to every term in the FROM clause. |
| 3551 | ** |
| 3552 | ** When assigning bitmask values to FROM clause cursors, it must be |
| 3553 | ** the case that if X is the bitmask for the N-th FROM clause term then |
| 3554 | ** the bitmask for all FROM clause terms to the left of the N-th term |
| 3555 | ** is (X-1). An expression from the ON clause of a LEFT JOIN can use |
| 3556 | ** its Expr.iRightJoinTable value to find the bitmask of the right table |
| 3557 | ** of the join. Subtracting one from the right table bitmask gives a |
| 3558 | ** bitmask for all tables to the left of the join. Knowing the bitmask |
| 3559 | ** for all tables to the left of a left join is important. Ticket #3015. |
danielk1977 | e672c8e | 2009-05-22 15:43:26 +0000 | [diff] [blame] | 3560 | ** |
| 3561 | ** Configure the WhereClause.vmask variable so that bits that correspond |
| 3562 | ** to virtual table cursors are set. This is used to selectively disable |
| 3563 | ** the OR-to-IN transformation in exprAnalyzeOrTerm(). It is not helpful |
| 3564 | ** with virtual tables. |
drh | 42165be | 2008-03-26 14:56:34 +0000 | [diff] [blame] | 3565 | */ |
danielk1977 | e672c8e | 2009-05-22 15:43:26 +0000 | [diff] [blame] | 3566 | assert( pWC->vmask==0 && pMaskSet->n==0 ); |
drh | 42165be | 2008-03-26 14:56:34 +0000 | [diff] [blame] | 3567 | for(i=0; i<pTabList->nSrc; i++){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3568 | createMask(pMaskSet, pTabList->a[i].iCursor); |
shane | e26fa4c | 2009-06-16 14:15:22 +0000 | [diff] [blame] | 3569 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
drh | 2c1a0c5 | 2009-06-11 17:04:28 +0000 | [diff] [blame] | 3570 | if( ALWAYS(pTabList->a[i].pTab) && IsVirtual(pTabList->a[i].pTab) ){ |
danielk1977 | e672c8e | 2009-05-22 15:43:26 +0000 | [diff] [blame] | 3571 | pWC->vmask |= ((Bitmask)1 << i); |
| 3572 | } |
shane | e26fa4c | 2009-06-16 14:15:22 +0000 | [diff] [blame] | 3573 | #endif |
drh | 42165be | 2008-03-26 14:56:34 +0000 | [diff] [blame] | 3574 | } |
| 3575 | #ifndef NDEBUG |
| 3576 | { |
| 3577 | Bitmask toTheLeft = 0; |
| 3578 | for(i=0; i<pTabList->nSrc; i++){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3579 | Bitmask m = getMask(pMaskSet, pTabList->a[i].iCursor); |
drh | 42165be | 2008-03-26 14:56:34 +0000 | [diff] [blame] | 3580 | assert( (m-1)==toTheLeft ); |
| 3581 | toTheLeft |= m; |
| 3582 | } |
| 3583 | } |
| 3584 | #endif |
| 3585 | |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3586 | /* Analyze all of the subexpressions. Note that exprAnalyze() might |
| 3587 | ** add new virtual terms onto the end of the WHERE clause. We do not |
| 3588 | ** want to analyze these virtual terms, so start analyzing at the end |
drh | b6fb62d | 2005-09-20 08:47:20 +0000 | [diff] [blame] | 3589 | ** and work forward so that the added virtual terms are never processed. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3590 | */ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3591 | exprAnalyzeAll(pTabList, pWC); |
drh | 1743575 | 2007-08-16 04:30:38 +0000 | [diff] [blame] | 3592 | if( db->mallocFailed ){ |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 3593 | goto whereBeginError; |
drh | 0bbaa1b | 2005-08-19 19:14:12 +0000 | [diff] [blame] | 3594 | } |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3595 | |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3596 | /* Chose the best index to use for each table in the FROM clause. |
| 3597 | ** |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 3598 | ** This loop fills in the following fields: |
| 3599 | ** |
| 3600 | ** pWInfo->a[].pIdx The index to use for this level of the loop. |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 3601 | ** pWInfo->a[].wsFlags WHERE_xxx flags associated with pIdx |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 3602 | ** pWInfo->a[].nEq The number of == and IN constraints |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 3603 | ** pWInfo->a[].iFrom Which term of the FROM clause is being coded |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 3604 | ** pWInfo->a[].iTabCur The VDBE cursor for the database table |
| 3605 | ** pWInfo->a[].iIdxCur The VDBE cursor for the index |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3606 | ** pWInfo->a[].pTerm When wsFlags==WO_OR, the OR-clause term |
drh | 51147ba | 2005-07-23 22:59:55 +0000 | [diff] [blame] | 3607 | ** |
| 3608 | ** This loop also figures out the nesting order of tables in the FROM |
| 3609 | ** clause. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3610 | */ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 3611 | notReady = ~(Bitmask)0; |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3612 | pTabItem = pTabList->a; |
| 3613 | pLevel = pWInfo->a; |
drh | 943af3c | 2005-07-29 19:43:58 +0000 | [diff] [blame] | 3614 | andFlags = ~0; |
drh | 4f0c587 | 2007-03-26 22:05:01 +0000 | [diff] [blame] | 3615 | WHERETRACE(("*** Optimizer Start ***\n")); |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3616 | for(i=iFrom=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3617 | WhereCost bestPlan; /* Most efficient plan seen so far */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3618 | Index *pIdx; /* Index for FROM table at pTabItem */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3619 | int j; /* For looping over FROM tables */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 3620 | int bestJ = -1; /* The value of j */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3621 | Bitmask m; /* Bitmask value for j or bestJ */ |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 3622 | int isOptimal; /* Iterator for optimal/non-optimal search */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3623 | |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3624 | memset(&bestPlan, 0, sizeof(bestPlan)); |
| 3625 | bestPlan.rCost = SQLITE_BIG_DBL; |
drh | df26fd5 | 2006-06-06 11:45:54 +0000 | [diff] [blame] | 3626 | |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 3627 | /* Loop through the remaining entries in the FROM clause to find the |
| 3628 | ** next nested loop. The FROM clause entries may be iterated through |
| 3629 | ** either once or twice. |
| 3630 | ** |
| 3631 | ** The first iteration, which is always performed, searches for the |
| 3632 | ** FROM clause entry that permits the lowest-cost, "optimal" scan. In |
| 3633 | ** this context an optimal scan is one that uses the same strategy |
| 3634 | ** for the given FROM clause entry as would be selected if the entry |
drh | d001516 | 2009-08-21 13:22:25 +0000 | [diff] [blame] | 3635 | ** were used as the innermost nested loop. In other words, a table |
| 3636 | ** is chosen such that the cost of running that table cannot be reduced |
| 3637 | ** by waiting for other tables to run first. |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 3638 | ** |
| 3639 | ** The second iteration is only performed if no optimal scan strategies |
| 3640 | ** were found by the first. This iteration is used to search for the |
| 3641 | ** lowest cost scan overall. |
| 3642 | ** |
| 3643 | ** Previous versions of SQLite performed only the second iteration - |
| 3644 | ** the next outermost loop was always that with the lowest overall |
| 3645 | ** cost. However, this meant that SQLite could select the wrong plan |
| 3646 | ** for scripts such as the following: |
| 3647 | ** |
| 3648 | ** CREATE TABLE t1(a, b); |
| 3649 | ** CREATE TABLE t2(c, d); |
| 3650 | ** SELECT * FROM t2, t1 WHERE t2.rowid = t1.a; |
| 3651 | ** |
| 3652 | ** The best strategy is to iterate through table t1 first. However it |
| 3653 | ** is not possible to determine this with a simple greedy algorithm. |
| 3654 | ** However, since the cost of a linear scan through table t2 is the same |
| 3655 | ** as the cost of a linear scan through table t1, a simple greedy |
| 3656 | ** algorithm may choose to use t2 for the outer loop, which is a much |
| 3657 | ** costlier approach. |
| 3658 | */ |
| 3659 | for(isOptimal=1; isOptimal>=0 && bestJ<0; isOptimal--){ |
| 3660 | Bitmask mask = (isOptimal ? 0 : notReady); |
| 3661 | assert( (pTabList->nSrc-iFrom)>1 || isOptimal ); |
| 3662 | for(j=iFrom, pTabItem=&pTabList->a[j]; j<pTabList->nSrc; j++, pTabItem++){ |
| 3663 | int doNotReorder; /* True if this table should not be reordered */ |
| 3664 | WhereCost sCost; /* Cost information from best[Virtual]Index() */ |
| 3665 | ExprList *pOrderBy; /* ORDER BY clause for index to optimize */ |
| 3666 | |
| 3667 | doNotReorder = (pTabItem->jointype & (JT_LEFT|JT_CROSS))!=0; |
| 3668 | if( j!=iFrom && doNotReorder ) break; |
| 3669 | m = getMask(pMaskSet, pTabItem->iCursor); |
| 3670 | if( (m & notReady)==0 ){ |
| 3671 | if( j==iFrom ) iFrom++; |
| 3672 | continue; |
| 3673 | } |
| 3674 | pOrderBy = ((i==0 && ppOrderBy )?*ppOrderBy:0); |
| 3675 | |
| 3676 | assert( pTabItem->pTab ); |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3677 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 3678 | if( IsVirtual(pTabItem->pTab) ){ |
| 3679 | sqlite3_index_info **pp = &pWInfo->a[j].pIdxInfo; |
| 3680 | bestVirtualIndex(pParse, pWC, pTabItem, mask, pOrderBy, &sCost, pp); |
| 3681 | }else |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3682 | #endif |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 3683 | { |
| 3684 | bestBtreeIndex(pParse, pWC, pTabItem, mask, pOrderBy, &sCost); |
| 3685 | } |
| 3686 | assert( isOptimal || (sCost.used¬Ready)==0 ); |
| 3687 | |
| 3688 | if( (sCost.used¬Ready)==0 |
| 3689 | && (j==iFrom || sCost.rCost<bestPlan.rCost) |
| 3690 | ){ |
| 3691 | bestPlan = sCost; |
| 3692 | bestJ = j; |
| 3693 | } |
| 3694 | if( doNotReorder ) break; |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3695 | } |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3696 | } |
dan | 5236ac1 | 2009-08-13 07:09:33 +0000 | [diff] [blame] | 3697 | assert( bestJ>=0 ); |
danielk1977 | 992347f | 2008-12-30 09:45:45 +0000 | [diff] [blame] | 3698 | assert( notReady & getMask(pMaskSet, pTabList->a[bestJ].iCursor) ); |
drh | cb04134 | 2008-06-12 00:07:29 +0000 | [diff] [blame] | 3699 | WHERETRACE(("*** Optimizer selects table %d for loop %d\n", bestJ, |
drh | 3dec223 | 2005-09-10 15:28:09 +0000 | [diff] [blame] | 3700 | pLevel-pWInfo->a)); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3701 | if( (bestPlan.plan.wsFlags & WHERE_ORDERBY)!=0 ){ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 3702 | *ppOrderBy = 0; |
drh | c4a3c77 | 2001-04-04 11:48:57 +0000 | [diff] [blame] | 3703 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3704 | andFlags &= bestPlan.plan.wsFlags; |
| 3705 | pLevel->plan = bestPlan.plan; |
| 3706 | if( bestPlan.plan.wsFlags & WHERE_INDEXED ){ |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3707 | pLevel->iIdxCur = pParse->nTab++; |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 3708 | }else{ |
| 3709 | pLevel->iIdxCur = -1; |
drh | 6b56344 | 2001-11-07 16:48:26 +0000 | [diff] [blame] | 3710 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3711 | notReady &= ~getMask(pMaskSet, pTabList->a[bestJ].iCursor); |
shane | d87897d | 2009-01-30 05:40:27 +0000 | [diff] [blame] | 3712 | pLevel->iFrom = (u8)bestJ; |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 3713 | |
| 3714 | /* Check that if the table scanned by this loop iteration had an |
| 3715 | ** INDEXED BY clause attached to it, that the named index is being |
| 3716 | ** used for the scan. If not, then query compilation has failed. |
| 3717 | ** Return an error. |
| 3718 | */ |
| 3719 | pIdx = pTabList->a[bestJ].pIndex; |
drh | 171256c | 2009-01-08 03:11:19 +0000 | [diff] [blame] | 3720 | if( pIdx ){ |
| 3721 | if( (bestPlan.plan.wsFlags & WHERE_INDEXED)==0 ){ |
| 3722 | sqlite3ErrorMsg(pParse, "cannot use index: %s", pIdx->zName); |
| 3723 | goto whereBeginError; |
| 3724 | }else{ |
| 3725 | /* If an INDEXED BY clause is used, the bestIndex() function is |
| 3726 | ** guaranteed to find the index specified in the INDEXED BY clause |
| 3727 | ** if it find an index at all. */ |
| 3728 | assert( bestPlan.plan.u.pIdx==pIdx ); |
| 3729 | } |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 3730 | } |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3731 | } |
drh | 4f0c587 | 2007-03-26 22:05:01 +0000 | [diff] [blame] | 3732 | WHERETRACE(("*** Optimizer Finished ***\n")); |
danielk1977 | 1d46146 | 2009-04-21 09:02:45 +0000 | [diff] [blame] | 3733 | if( pParse->nErr || db->mallocFailed ){ |
danielk1977 | 8044294 | 2008-12-24 11:25:39 +0000 | [diff] [blame] | 3734 | goto whereBeginError; |
| 3735 | } |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3736 | |
drh | 943af3c | 2005-07-29 19:43:58 +0000 | [diff] [blame] | 3737 | /* If the total query only selects a single row, then the ORDER BY |
| 3738 | ** clause is irrelevant. |
| 3739 | */ |
| 3740 | if( (andFlags & WHERE_UNIQUE)!=0 && ppOrderBy ){ |
| 3741 | *ppOrderBy = 0; |
| 3742 | } |
| 3743 | |
drh | 08c88eb | 2008-04-10 13:33:18 +0000 | [diff] [blame] | 3744 | /* If the caller is an UPDATE or DELETE statement that is requesting |
| 3745 | ** to use a one-pass algorithm, determine if this is appropriate. |
| 3746 | ** The one-pass algorithm only works if the WHERE clause constraints |
| 3747 | ** the statement to update a single row. |
| 3748 | */ |
drh | 165be38 | 2008-12-05 02:36:33 +0000 | [diff] [blame] | 3749 | assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); |
| 3750 | if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 && (andFlags & WHERE_UNIQUE)!=0 ){ |
drh | 08c88eb | 2008-04-10 13:33:18 +0000 | [diff] [blame] | 3751 | pWInfo->okOnePass = 1; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3752 | pWInfo->a[0].plan.wsFlags &= ~WHERE_IDX_ONLY; |
drh | 08c88eb | 2008-04-10 13:33:18 +0000 | [diff] [blame] | 3753 | } |
| 3754 | |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3755 | /* Open all tables in the pTabList and any indices selected for |
| 3756 | ** searching those tables. |
| 3757 | */ |
| 3758 | sqlite3CodeVerifySchema(pParse, -1); /* Insert the cookie verifier Goto */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3759 | for(i=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){ |
danielk1977 | da18423 | 2006-01-05 11:34:32 +0000 | [diff] [blame] | 3760 | Table *pTab; /* Table to open */ |
danielk1977 | da18423 | 2006-01-05 11:34:32 +0000 | [diff] [blame] | 3761 | int iDb; /* Index of database containing table/index */ |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3762 | |
drh | ecc9242 | 2005-09-10 16:46:12 +0000 | [diff] [blame] | 3763 | #ifndef SQLITE_OMIT_EXPLAIN |
| 3764 | if( pParse->explain==2 ){ |
| 3765 | char *zMsg; |
| 3766 | struct SrcList_item *pItem = &pTabList->a[pLevel->iFrom]; |
danielk1977 | 1e53695 | 2007-08-16 10:09:01 +0000 | [diff] [blame] | 3767 | zMsg = sqlite3MPrintf(db, "TABLE %s", pItem->zName); |
drh | ecc9242 | 2005-09-10 16:46:12 +0000 | [diff] [blame] | 3768 | if( pItem->zAlias ){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 3769 | zMsg = sqlite3MAppendf(db, zMsg, "%s AS %s", zMsg, pItem->zAlias); |
drh | ecc9242 | 2005-09-10 16:46:12 +0000 | [diff] [blame] | 3770 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3771 | if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){ |
| 3772 | zMsg = sqlite3MAppendf(db, zMsg, "%s WITH INDEX %s", |
| 3773 | zMsg, pLevel->plan.u.pIdx->zName); |
drh | 46129af | 2008-12-30 16:18:47 +0000 | [diff] [blame] | 3774 | }else if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){ |
| 3775 | zMsg = sqlite3MAppendf(db, zMsg, "%s VIA MULTI-INDEX UNION", zMsg); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3776 | }else if( pLevel->plan.wsFlags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 3777 | zMsg = sqlite3MAppendf(db, zMsg, "%s USING PRIMARY KEY", zMsg); |
drh | ecc9242 | 2005-09-10 16:46:12 +0000 | [diff] [blame] | 3778 | } |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3779 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3780 | else if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){ |
| 3781 | sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx; |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 3782 | zMsg = sqlite3MAppendf(db, zMsg, "%s VIRTUAL TABLE INDEX %d:%s", zMsg, |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3783 | pVtabIdx->idxNum, pVtabIdx->idxStr); |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3784 | } |
| 3785 | #endif |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3786 | if( pLevel->plan.wsFlags & WHERE_ORDERBY ){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 3787 | zMsg = sqlite3MAppendf(db, zMsg, "%s ORDER BY", zMsg); |
drh | e2b3909 | 2006-04-21 09:38:36 +0000 | [diff] [blame] | 3788 | } |
drh | 66a5167 | 2008-01-03 00:01:23 +0000 | [diff] [blame] | 3789 | sqlite3VdbeAddOp4(v, OP_Explain, i, pLevel->iFrom, 0, zMsg, P4_DYNAMIC); |
drh | ecc9242 | 2005-09-10 16:46:12 +0000 | [diff] [blame] | 3790 | } |
| 3791 | #endif /* SQLITE_OMIT_EXPLAIN */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3792 | pTabItem = &pTabList->a[pLevel->iFrom]; |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3793 | pTab = pTabItem->pTab; |
danielk1977 | 595a523 | 2009-07-24 17:58:53 +0000 | [diff] [blame] | 3794 | iDb = sqlite3SchemaToIndex(db, pTab->pSchema); |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 3795 | if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ) continue; |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3796 | #ifndef SQLITE_OMIT_VIRTUALTABLE |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3797 | if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){ |
danielk1977 | 595a523 | 2009-07-24 17:58:53 +0000 | [diff] [blame] | 3798 | const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); |
danielk1977 | 93626f4 | 2006-06-20 13:07:27 +0000 | [diff] [blame] | 3799 | int iCur = pTabItem->iCursor; |
danielk1977 | 595a523 | 2009-07-24 17:58:53 +0000 | [diff] [blame] | 3800 | sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); |
drh | 9eff616 | 2006-06-12 21:59:13 +0000 | [diff] [blame] | 3801 | }else |
| 3802 | #endif |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3803 | if( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 |
| 3804 | && (wctrlFlags & WHERE_OMIT_OPEN)==0 ){ |
drh | 08c88eb | 2008-04-10 13:33:18 +0000 | [diff] [blame] | 3805 | int op = pWInfo->okOnePass ? OP_OpenWrite : OP_OpenRead; |
| 3806 | sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); |
danielk1977 | 2343297 | 2008-11-17 16:42:00 +0000 | [diff] [blame] | 3807 | if( !pWInfo->okOnePass && pTab->nCol<BMS ){ |
danielk1977 | 9792eef | 2006-01-13 15:58:43 +0000 | [diff] [blame] | 3808 | Bitmask b = pTabItem->colUsed; |
| 3809 | int n = 0; |
drh | 7416170 | 2006-02-24 02:53:49 +0000 | [diff] [blame] | 3810 | for(; b; b=b>>1, n++){} |
shane | c0688ea | 2009-03-05 03:48:06 +0000 | [diff] [blame] | 3811 | sqlite3VdbeChangeP4(v, sqlite3VdbeCurrentAddr(v)-1, SQLITE_INT_TO_PTR(n), P4_INT32); |
danielk1977 | 9792eef | 2006-01-13 15:58:43 +0000 | [diff] [blame] | 3812 | assert( n<=pTab->nCol ); |
| 3813 | } |
danielk1977 | c00da10 | 2006-01-07 13:21:04 +0000 | [diff] [blame] | 3814 | }else{ |
| 3815 | sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3816 | } |
| 3817 | pLevel->iTabCur = pTabItem->iCursor; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3818 | if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){ |
| 3819 | Index *pIx = pLevel->plan.u.pIdx; |
danielk1977 | b3bf556 | 2006-01-10 17:58:23 +0000 | [diff] [blame] | 3820 | KeyInfo *pKey = sqlite3IndexKeyinfo(pParse, pIx); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3821 | int iIdxCur = pLevel->iIdxCur; |
danielk1977 | da18423 | 2006-01-05 11:34:32 +0000 | [diff] [blame] | 3822 | assert( pIx->pSchema==pTab->pSchema ); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3823 | assert( iIdxCur>=0 ); |
danielk1977 | 207872a | 2008-01-03 07:54:23 +0000 | [diff] [blame] | 3824 | sqlite3VdbeAddOp4(v, OP_OpenRead, iIdxCur, pIx->tnum, iDb, |
drh | 66a5167 | 2008-01-03 00:01:23 +0000 | [diff] [blame] | 3825 | (char*)pKey, P4_KEYINFO_HANDOFF); |
danielk1977 | 207872a | 2008-01-03 07:54:23 +0000 | [diff] [blame] | 3826 | VdbeComment((v, "%s", pIx->zName)); |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3827 | } |
danielk1977 | da18423 | 2006-01-05 11:34:32 +0000 | [diff] [blame] | 3828 | sqlite3CodeVerifySchema(pParse, iDb); |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3829 | } |
| 3830 | pWInfo->iTop = sqlite3VdbeCurrentAddr(v); |
| 3831 | |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3832 | /* Generate the code to do the search. Each iteration of the for |
| 3833 | ** loop below generates code for a single nested loop of the VM |
| 3834 | ** program. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3835 | */ |
drh | fe05af8 | 2005-07-21 03:14:59 +0000 | [diff] [blame] | 3836 | notReady = ~(Bitmask)0; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3837 | for(i=0; i<pTabList->nSrc; i++){ |
| 3838 | notReady = codeOneLoopStart(pWInfo, i, wctrlFlags, notReady); |
drh | 813f31e | 2009-01-06 00:08:02 +0000 | [diff] [blame] | 3839 | pWInfo->iContinue = pWInfo->a[i].addrCont; |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3840 | } |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3841 | |
| 3842 | #ifdef SQLITE_TEST /* For testing and debugging use only */ |
| 3843 | /* Record in the query plan information about the current table |
| 3844 | ** and the index used to access it (if any). If the table itself |
| 3845 | ** is not used, its name is just '{}'. If no index is used |
| 3846 | ** the index is listed as "{}". If the primary key is used the |
| 3847 | ** index name is '*'. |
| 3848 | */ |
| 3849 | for(i=0; i<pTabList->nSrc; i++){ |
| 3850 | char *z; |
| 3851 | int n; |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3852 | pLevel = &pWInfo->a[i]; |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3853 | pTabItem = &pTabList->a[pLevel->iFrom]; |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3854 | z = pTabItem->zAlias; |
| 3855 | if( z==0 ) z = pTabItem->pTab->zName; |
drh | ea67883 | 2008-12-10 19:26:22 +0000 | [diff] [blame] | 3856 | n = sqlite3Strlen30(z); |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3857 | if( n+nQPlan < sizeof(sqlite3_query_plan)-10 ){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3858 | if( pLevel->plan.wsFlags & WHERE_IDX_ONLY ){ |
drh | 5bb3eb9 | 2007-05-04 13:15:55 +0000 | [diff] [blame] | 3859 | memcpy(&sqlite3_query_plan[nQPlan], "{}", 2); |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3860 | nQPlan += 2; |
| 3861 | }else{ |
drh | 5bb3eb9 | 2007-05-04 13:15:55 +0000 | [diff] [blame] | 3862 | memcpy(&sqlite3_query_plan[nQPlan], z, n); |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3863 | nQPlan += n; |
| 3864 | } |
| 3865 | sqlite3_query_plan[nQPlan++] = ' '; |
| 3866 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3867 | testcase( pLevel->plan.wsFlags & WHERE_ROWID_EQ ); |
| 3868 | testcase( pLevel->plan.wsFlags & WHERE_ROWID_RANGE ); |
| 3869 | if( pLevel->plan.wsFlags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){ |
drh | 5bb3eb9 | 2007-05-04 13:15:55 +0000 | [diff] [blame] | 3870 | memcpy(&sqlite3_query_plan[nQPlan], "* ", 2); |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3871 | nQPlan += 2; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3872 | }else if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){ |
| 3873 | n = sqlite3Strlen30(pLevel->plan.u.pIdx->zName); |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3874 | if( n+nQPlan < sizeof(sqlite3_query_plan)-2 ){ |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3875 | memcpy(&sqlite3_query_plan[nQPlan], pLevel->plan.u.pIdx->zName, n); |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3876 | nQPlan += n; |
| 3877 | sqlite3_query_plan[nQPlan++] = ' '; |
| 3878 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3879 | }else{ |
| 3880 | memcpy(&sqlite3_query_plan[nQPlan], "{} ", 3); |
| 3881 | nQPlan += 3; |
drh | 7ec764a | 2005-07-21 03:48:20 +0000 | [diff] [blame] | 3882 | } |
| 3883 | } |
| 3884 | while( nQPlan>0 && sqlite3_query_plan[nQPlan-1]==' ' ){ |
| 3885 | sqlite3_query_plan[--nQPlan] = 0; |
| 3886 | } |
| 3887 | sqlite3_query_plan[nQPlan] = 0; |
| 3888 | nQPlan = 0; |
| 3889 | #endif /* SQLITE_TEST // Testing and debugging use only */ |
| 3890 | |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3891 | /* Record the continuation address in the WhereInfo structure. Then |
| 3892 | ** clean up and return. |
| 3893 | */ |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3894 | return pWInfo; |
drh | e23399f | 2005-07-22 00:31:39 +0000 | [diff] [blame] | 3895 | |
| 3896 | /* Jump here if malloc fails */ |
danielk1977 | 85574e3 | 2008-10-06 05:32:18 +0000 | [diff] [blame] | 3897 | whereBeginError: |
drh | 10fe840 | 2008-10-11 16:47:35 +0000 | [diff] [blame] | 3898 | whereInfoFree(db, pWInfo); |
drh | e23399f | 2005-07-22 00:31:39 +0000 | [diff] [blame] | 3899 | return 0; |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3900 | } |
| 3901 | |
| 3902 | /* |
drh | c27a1ce | 2002-06-14 20:58:45 +0000 | [diff] [blame] | 3903 | ** Generate the end of the WHERE loop. See comments on |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3904 | ** sqlite3WhereBegin() for additional information. |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 3905 | */ |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3906 | void sqlite3WhereEnd(WhereInfo *pWInfo){ |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 3907 | Parse *pParse = pWInfo->pParse; |
| 3908 | Vdbe *v = pParse->pVdbe; |
drh | 19a775c | 2000-06-05 18:54:46 +0000 | [diff] [blame] | 3909 | int i; |
drh | 6b56344 | 2001-11-07 16:48:26 +0000 | [diff] [blame] | 3910 | WhereLevel *pLevel; |
drh | ad3cab5 | 2002-05-24 02:04:32 +0000 | [diff] [blame] | 3911 | SrcList *pTabList = pWInfo->pTabList; |
drh | 633e6d5 | 2008-07-28 19:34:53 +0000 | [diff] [blame] | 3912 | sqlite3 *db = pParse->db; |
drh | 19a775c | 2000-06-05 18:54:46 +0000 | [diff] [blame] | 3913 | |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3914 | /* Generate loop termination code. |
| 3915 | */ |
drh | ceea332 | 2009-04-23 13:22:42 +0000 | [diff] [blame] | 3916 | sqlite3ExprCacheClear(pParse); |
drh | ad3cab5 | 2002-05-24 02:04:32 +0000 | [diff] [blame] | 3917 | for(i=pTabList->nSrc-1; i>=0; i--){ |
drh | 6b56344 | 2001-11-07 16:48:26 +0000 | [diff] [blame] | 3918 | pLevel = &pWInfo->a[i]; |
drh | b3190c1 | 2008-12-08 21:37:14 +0000 | [diff] [blame] | 3919 | sqlite3VdbeResolveLabel(v, pLevel->addrCont); |
drh | 6b56344 | 2001-11-07 16:48:26 +0000 | [diff] [blame] | 3920 | if( pLevel->op!=OP_Noop ){ |
drh | 66a5167 | 2008-01-03 00:01:23 +0000 | [diff] [blame] | 3921 | sqlite3VdbeAddOp2(v, pLevel->op, pLevel->p1, pLevel->p2); |
drh | d1d3848 | 2008-10-07 23:46:38 +0000 | [diff] [blame] | 3922 | sqlite3VdbeChangeP5(v, pLevel->p5); |
drh | 19a775c | 2000-06-05 18:54:46 +0000 | [diff] [blame] | 3923 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3924 | if( pLevel->plan.wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ |
drh | 72e8fa4 | 2007-03-28 14:30:06 +0000 | [diff] [blame] | 3925 | struct InLoop *pIn; |
drh | e23399f | 2005-07-22 00:31:39 +0000 | [diff] [blame] | 3926 | int j; |
drh | b3190c1 | 2008-12-08 21:37:14 +0000 | [diff] [blame] | 3927 | sqlite3VdbeResolveLabel(v, pLevel->addrNxt); |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3928 | for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ |
drh | b3190c1 | 2008-12-08 21:37:14 +0000 | [diff] [blame] | 3929 | sqlite3VdbeJumpHere(v, pIn->addrInTop+1); |
| 3930 | sqlite3VdbeAddOp2(v, OP_Next, pIn->iCur, pIn->addrInTop); |
| 3931 | sqlite3VdbeJumpHere(v, pIn->addrInTop-1); |
drh | e23399f | 2005-07-22 00:31:39 +0000 | [diff] [blame] | 3932 | } |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3933 | sqlite3DbFree(db, pLevel->u.in.aInLoop); |
drh | d99f706 | 2002-06-08 23:25:08 +0000 | [diff] [blame] | 3934 | } |
drh | b3190c1 | 2008-12-08 21:37:14 +0000 | [diff] [blame] | 3935 | sqlite3VdbeResolveLabel(v, pLevel->addrBrk); |
drh | ad2d830 | 2002-05-24 20:31:36 +0000 | [diff] [blame] | 3936 | if( pLevel->iLeftJoin ){ |
| 3937 | int addr; |
drh | 3c84ddf | 2008-01-09 02:15:38 +0000 | [diff] [blame] | 3938 | addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); |
| 3939 | sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor); |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3940 | if( pLevel->iIdxCur>=0 ){ |
drh | 3c84ddf | 2008-01-09 02:15:38 +0000 | [diff] [blame] | 3941 | sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); |
drh | 7f09b3e | 2002-08-13 13:15:49 +0000 | [diff] [blame] | 3942 | } |
drh | 336a530 | 2009-04-24 15:46:21 +0000 | [diff] [blame] | 3943 | if( pLevel->op==OP_Return ){ |
| 3944 | sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); |
| 3945 | }else{ |
| 3946 | sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrFirst); |
| 3947 | } |
drh | d654be8 | 2005-09-20 17:42:23 +0000 | [diff] [blame] | 3948 | sqlite3VdbeJumpHere(v, addr); |
drh | ad2d830 | 2002-05-24 20:31:36 +0000 | [diff] [blame] | 3949 | } |
drh | 19a775c | 2000-06-05 18:54:46 +0000 | [diff] [blame] | 3950 | } |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3951 | |
| 3952 | /* The "break" point is here, just past the end of the outer loop. |
| 3953 | ** Set it. |
| 3954 | */ |
danielk1977 | 4adee20 | 2004-05-08 08:23:19 +0000 | [diff] [blame] | 3955 | sqlite3VdbeResolveLabel(v, pWInfo->iBreak); |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3956 | |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3957 | /* Close all of the cursors that were opened by sqlite3WhereBegin. |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3958 | */ |
drh | 29dda4a | 2005-07-21 18:23:20 +0000 | [diff] [blame] | 3959 | for(i=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){ |
| 3960 | struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3961 | Table *pTab = pTabItem->pTab; |
drh | 5cf590c | 2003-04-24 01:45:04 +0000 | [diff] [blame] | 3962 | assert( pTab!=0 ); |
drh | 7d10d5a | 2008-08-20 16:35:10 +0000 | [diff] [blame] | 3963 | if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ) continue; |
drh | 6df2acd | 2008-12-28 16:55:25 +0000 | [diff] [blame] | 3964 | if( (pWInfo->wctrlFlags & WHERE_OMIT_CLOSE)==0 ){ |
| 3965 | if( !pWInfo->okOnePass && (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 ){ |
| 3966 | sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); |
| 3967 | } |
| 3968 | if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){ |
| 3969 | sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); |
| 3970 | } |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3971 | } |
| 3972 | |
danielk1977 | 21de2e7 | 2007-11-29 17:43:27 +0000 | [diff] [blame] | 3973 | /* If this scan uses an index, make code substitutions to read data |
| 3974 | ** from the index in preference to the table. Sometimes, this means |
| 3975 | ** the table need never be read from. This is a performance boost, |
| 3976 | ** as the vdbe level waits until the table is read before actually |
| 3977 | ** seeking the table cursor to the record corresponding to the current |
| 3978 | ** position in the index. |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3979 | ** |
| 3980 | ** Calls to the code generator in between sqlite3WhereBegin and |
| 3981 | ** sqlite3WhereEnd will have created code that references the table |
| 3982 | ** directly. This loop scans all that code looking for opcodes |
| 3983 | ** that reference the table and converts them into opcodes that |
| 3984 | ** reference the index. |
| 3985 | */ |
drh | 125feff | 2009-06-06 15:17:27 +0000 | [diff] [blame] | 3986 | if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 && !db->mallocFailed){ |
danielk1977 | f011300 | 2006-01-24 12:09:17 +0000 | [diff] [blame] | 3987 | int k, j, last; |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3988 | VdbeOp *pOp; |
drh | 111a6a7 | 2008-12-21 03:51:16 +0000 | [diff] [blame] | 3989 | Index *pIdx = pLevel->plan.u.pIdx; |
| 3990 | int useIndexOnly = pLevel->plan.wsFlags & WHERE_IDX_ONLY; |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3991 | |
| 3992 | assert( pIdx!=0 ); |
| 3993 | pOp = sqlite3VdbeGetOp(v, pWInfo->iTop); |
| 3994 | last = sqlite3VdbeCurrentAddr(v); |
danielk1977 | f011300 | 2006-01-24 12:09:17 +0000 | [diff] [blame] | 3995 | for(k=pWInfo->iTop; k<last; k++, pOp++){ |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3996 | if( pOp->p1!=pLevel->iTabCur ) continue; |
| 3997 | if( pOp->opcode==OP_Column ){ |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 3998 | for(j=0; j<pIdx->nColumn; j++){ |
| 3999 | if( pOp->p2==pIdx->aiColumn[j] ){ |
| 4000 | pOp->p2 = j; |
danielk1977 | 21de2e7 | 2007-11-29 17:43:27 +0000 | [diff] [blame] | 4001 | pOp->p1 = pLevel->iIdxCur; |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 4002 | break; |
| 4003 | } |
| 4004 | } |
danielk1977 | 21de2e7 | 2007-11-29 17:43:27 +0000 | [diff] [blame] | 4005 | assert(!useIndexOnly || j<pIdx->nColumn); |
drh | f0863fe | 2005-06-12 21:35:51 +0000 | [diff] [blame] | 4006 | }else if( pOp->opcode==OP_Rowid ){ |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 4007 | pOp->p1 = pLevel->iIdxCur; |
drh | f0863fe | 2005-06-12 21:35:51 +0000 | [diff] [blame] | 4008 | pOp->opcode = OP_IdxRowid; |
danielk1977 | 21de2e7 | 2007-11-29 17:43:27 +0000 | [diff] [blame] | 4009 | }else if( pOp->opcode==OP_NullRow && useIndexOnly ){ |
danielk1977 | 6c18b6e | 2005-01-30 09:17:58 +0000 | [diff] [blame] | 4010 | pOp->opcode = OP_Noop; |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 4011 | } |
| 4012 | } |
drh | 6b56344 | 2001-11-07 16:48:26 +0000 | [diff] [blame] | 4013 | } |
drh | 19a775c | 2000-06-05 18:54:46 +0000 | [diff] [blame] | 4014 | } |
drh | 9012bcb | 2004-12-19 00:11:35 +0000 | [diff] [blame] | 4015 | |
| 4016 | /* Final cleanup |
| 4017 | */ |
drh | 10fe840 | 2008-10-11 16:47:35 +0000 | [diff] [blame] | 4018 | whereInfoFree(db, pWInfo); |
drh | 7589723 | 2000-05-29 14:26:00 +0000 | [diff] [blame] | 4019 | return; |
| 4020 | } |