Remove the Expr.pColl field and compute the collating sequence as it is needed.
This fixes the test script "shared9.test", though there is still a memory leak.
And there are other problems. Consider this a work-in-progress.
FossilOrigin-Name: fd011cb22f2d899d94ec7ce22641d7a8f5e19972
diff --git a/src/build.c b/src/build.c
index 4c45a92..c03e502 100644
--- a/src/build.c
+++ b/src/build.c
@@ -2690,10 +2690,8 @@
for(i=0; i<pList->nExpr; i++){
Expr *pExpr = pList->a[i].pExpr;
if( pExpr ){
- CollSeq *pColl = pExpr->pColl;
- /* Either pColl!=0 or there was an OOM failure. But if an OOM
- ** failure we have quit before reaching this point. */
- if( ALWAYS(pColl) ){
+ CollSeq *pColl = sqlite3ExprCollSeq(pParse, pExpr);
+ if( pColl ){
nExtra += (1 + sqlite3Strlen30(pColl->zName));
}
}
@@ -2756,6 +2754,7 @@
const char *zColName = pListItem->zName;
Column *pTabCol;
int requestedSortOrder;
+ CollSeq *pColl; /* Collating sequence */
char *zColl; /* Collation sequence name */
for(j=0, pTabCol=pTab->aCol; j<pTab->nCol; j++, pTabCol++){
@@ -2768,14 +2767,11 @@
goto exit_create_index;
}
pIndex->aiColumn[i] = j;
- /* Justification of the ALWAYS(pListItem->pExpr->pColl): Because of
- ** the way the "idxlist" non-terminal is constructed by the parser,
- ** if pListItem->pExpr is not null then either pListItem->pExpr->pColl
- ** must exist or else there must have been an OOM error. But if there
- ** was an OOM error, we would never reach this point. */
- if( pListItem->pExpr && ALWAYS(pListItem->pExpr->pColl) ){
+ if( pListItem->pExpr
+ && (pColl = sqlite3ExprCollSeq(pParse, pListItem->pExpr))!=0
+ ){
int nColl;
- zColl = pListItem->pExpr->pColl->zName;
+ zColl = pColl->zName;
nColl = sqlite3Strlen30(zColl) + 1;
assert( nExtra>=nColl );
memcpy(zExtra, zColl, nColl);
diff --git a/src/expr.c b/src/expr.c
index ae82848..11feaa2 100644
--- a/src/expr.c
+++ b/src/expr.c
@@ -56,66 +56,72 @@
}
/*
-** Set the explicit collating sequence for an expression to the
-** collating sequence supplied in the second argument.
-*/
-Expr *sqlite3ExprSetColl(Expr *pExpr, CollSeq *pColl){
- if( pExpr && pColl ){
- pExpr->pColl = pColl;
- pExpr->flags |= EP_ExpCollate;
- }
- return pExpr;
-}
-
-/*
** Set the collating sequence for expression pExpr to be the collating
-** sequence named by pToken. Return a pointer to the revised expression.
-** The collating sequence is marked as "explicit" using the EP_ExpCollate
-** flag. An explicit collating sequence will override implicit
-** collating sequences.
+** sequence named by pToken. Return a pointer to a new Expr node that
+** implements the COLLATE operator.
*/
Expr *sqlite3ExprSetCollByToken(Parse *pParse, Expr *pExpr, Token *pCollName){
- char *zColl = 0; /* Dequoted name of collation sequence */
- CollSeq *pColl;
- sqlite3 *db = pParse->db;
- zColl = sqlite3NameFromToken(db, pCollName);
- pColl = sqlite3LocateCollSeq(pParse, zColl);
- sqlite3ExprSetColl(pExpr, pColl);
- sqlite3DbFree(db, zColl);
- return pExpr;
+ Expr *pNew = sqlite3ExprAlloc(pParse->db, TK_COLLATE, pCollName, 1);
+ if( pNew ){
+ pNew->pLeft = pExpr;
+ pNew->flags |= EP_Collate;
+ }
+ return pNew;
}
/*
-** Return the default collation sequence for the expression pExpr. If
-** there is no default collation type, return 0.
+** Return the collation sequence for the expression pExpr. If
+** there is no defined collating sequence, return NULL.
+**
+** The collating sequence might be determined by a COLLATE operator
+** or by the presence of a column with a defined collating sequence.
+** COLLATE operators take first precedence. Left operands take
+** precedence over right operands.
*/
CollSeq *sqlite3ExprCollSeq(Parse *pParse, Expr *pExpr){
+ sqlite3 *db = pParse->db;
CollSeq *pColl = 0;
Expr *p = pExpr;
- while( p ){
- int op;
- pColl = p->pColl;
- if( pColl ) break;
- op = p->op;
- if( p->pTab!=0 && (
- op==TK_AGG_COLUMN || op==TK_COLUMN || op==TK_REGISTER || op==TK_TRIGGER
- )){
+ while( p && pColl==0 ){
+ int op = p->op;
+ if( op==TK_CAST || op==TK_UPLUS ){
+ p = p->pLeft;
+ continue;
+ }
+ if( op==TK_COLLATE ){
+ pColl = sqlite3FindCollSeq(db, ENC(db), p->u.zToken, 0);
+ break;
+ }
+ if( p->pTab!=0
+ && (op==TK_AGG_COLUMN || op==TK_COLUMN
+ || op==TK_REGISTER || op==TK_TRIGGER)
+ ){
/* op==TK_REGISTER && p->pTab!=0 happens when pExpr was originally
** a TK_COLUMN but was previously evaluated and cached in a register */
- const char *zColl;
int j = p->iColumn;
if( j>=0 ){
- sqlite3 *db = pParse->db;
- zColl = p->pTab->aCol[j].zColl;
+ const char *zColl = p->pTab->aCol[j].zColl;
pColl = sqlite3FindCollSeq(db, ENC(db), zColl, 0);
- pExpr->pColl = pColl;
}
break;
}
- if( op!=TK_CAST && op!=TK_UPLUS ){
+ if( p->flags & EP_Collate ){
+ if( p->pLeft->flags & EP_Collate ){
+ p = p->pLeft;
+ }else{
+ p = p->pRight;
+ }
+ }else{
break;
}
- p = p->pLeft;
+#if 0
+ else if( p->flags & EP_TokenOnly ){
+ break;
+ }else{
+ pColl = sqlite3ExprCollSeq(pParse, p->pLeft);
+ p = p->pRight;
+ }
+#endif
}
if( sqlite3CheckCollSeq(pParse, pColl) ){
pColl = 0;
@@ -219,12 +225,10 @@
){
CollSeq *pColl;
assert( pLeft );
- if( pLeft->flags & EP_ExpCollate ){
- assert( pLeft->pColl );
- pColl = pLeft->pColl;
- }else if( pRight && pRight->flags & EP_ExpCollate ){
- assert( pRight->pColl );
- pColl = pRight->pColl;
+ if( pLeft->flags & EP_Collate ){
+ pColl = sqlite3ExprCollSeq(pParse, pLeft);
+ }else if( pRight && (pRight->flags & EP_Collate)!=0 ){
+ pColl = sqlite3ExprCollSeq(pParse, pRight);
}else{
pColl = sqlite3ExprCollSeq(pParse, pLeft);
if( !pColl ){
@@ -454,17 +458,11 @@
}else{
if( pRight ){
pRoot->pRight = pRight;
- if( pRight->flags & EP_ExpCollate ){
- pRoot->flags |= EP_ExpCollate;
- pRoot->pColl = pRight->pColl;
- }
+ pRoot->flags |= EP_Collate & pRight->flags;
}
if( pLeft ){
pRoot->pLeft = pLeft;
- if( pLeft->flags & EP_ExpCollate ){
- pRoot->flags |= EP_ExpCollate;
- pRoot->pColl = pLeft->pColl;
- }
+ pRoot->flags |= EP_Collate & pLeft->flags;
}
exprSetHeight(pRoot);
}
@@ -722,7 +720,7 @@
assert( !ExprHasProperty(p, EP_FromJoin) );
assert( (p->flags2 & EP2_MallocedToken)==0 );
assert( (p->flags2 & EP2_Irreducible)==0 );
- if( p->pLeft || p->pRight || p->pColl || p->x.pList ){
+ if( p->pLeft || p->pRight || p->x.pList ){
nSize = EXPR_REDUCEDSIZE | EP_Reduced;
}else{
nSize = EXPR_TOKENONLYSIZE | EP_TokenOnly;
@@ -2746,6 +2744,7 @@
sqlite3ReleaseTempReg(pParse, r4);
break;
}
+ case TK_COLLATE:
case TK_UPLUS: {
inReg = sqlite3ExprCodeTarget(pParse, pExpr->pLeft, target);
break;
@@ -3773,7 +3772,15 @@
return 2;
}
if( (pA->flags & EP_Distinct)!=(pB->flags & EP_Distinct) ) return 2;
- if( pA->op!=pB->op ) return 2;
+ if( pA->op!=pB->op ){
+ if( pA->op==TK_COLLATE && sqlite3ExprCompare(pA->pLeft, pB)<2 ){
+ return 1;
+ }
+ if( pB->op==TK_COLLATE && sqlite3ExprCompare(pA, pB->pLeft)<2 ){
+ return 1;
+ }
+ return 2;
+ }
if( sqlite3ExprCompare(pA->pLeft, pB->pLeft) ) return 2;
if( sqlite3ExprCompare(pA->pRight, pB->pRight) ) return 2;
if( sqlite3ExprListCompare(pA->x.pList, pB->x.pList) ) return 2;
@@ -3785,11 +3792,9 @@
}else if( pA->op!=TK_COLUMN && ALWAYS(pA->op!=TK_AGG_COLUMN) && pA->u.zToken){
if( ExprHasProperty(pB, EP_IntValue) || NEVER(pB->u.zToken==0) ) return 2;
if( strcmp(pA->u.zToken,pB->u.zToken)!=0 ){
- return 2;
+ return pA->op==TK_COLLATE ? 1 : 2;
}
}
- if( (pA->flags & EP_ExpCollate)!=(pB->flags & EP_ExpCollate) ) return 1;
- if( (pA->flags & EP_ExpCollate)!=0 && pA->pColl!=pB->pColl ) return 2;
return 0;
}
diff --git a/src/fkey.c b/src/fkey.c
index a3214a9..a120b7b 100644
--- a/src/fkey.c
+++ b/src/fkey.c
@@ -516,7 +516,7 @@
if( pTab->iPKey==iCol ) iCol = -1;
pLeft->iTable = regData+iCol+1;
pLeft->affinity = pCol->affinity;
- pLeft->pColl = sqlite3LocateCollSeq(pParse, pCol->zColl);
+ // fix me. pLeft->pColl = sqlite3LocateCollSeq(pParse, pCol->zColl);
}else{
pLeft->iTable = regData;
pLeft->affinity = SQLITE_AFF_INTEGER;
diff --git a/src/resolve.c b/src/resolve.c
index 0ffc633..c2b0523 100644
--- a/src/resolve.c
+++ b/src/resolve.c
@@ -114,10 +114,6 @@
pDup->flags2 |= EP2_MallocedToken;
pDup->u.zToken = sqlite3DbStrDup(db, zToken);
}
- if( pExpr->flags & EP_ExpCollate ){
- pDup->pColl = pExpr->pColl;
- pDup->flags |= EP_ExpCollate;
- }
/* Before calling sqlite3ExprDelete(), set the EP_Static flag. This
** prevents ExprDelete() from deleting the Expr structure itself,
@@ -830,13 +826,10 @@
}
}
if( iCol>0 ){
- CollSeq *pColl = pE->pColl;
- int flags = pE->flags & EP_ExpCollate;
sqlite3ExprDelete(db, pE);
pItem->pExpr = pE = sqlite3Expr(db, TK_INTEGER, 0);
if( pE==0 ) return 1;
- pE->pColl = pColl;
- pE->flags |= EP_IntValue | flags;
+ pE->flags |= EP_IntValue;
pE->u.iValue = iCol;
pItem->iOrderByCol = (u16)iCol;
pItem->done = 1;
diff --git a/src/select.c b/src/select.c
index cba766f..ee1755d 100644
--- a/src/select.c
+++ b/src/select.c
@@ -2333,12 +2333,10 @@
for(i=0; i<nOrderBy; i++){
CollSeq *pColl;
Expr *pTerm = pOrderBy->a[i].pExpr;
- if( pTerm->flags & EP_ExpCollate ){
- pColl = pTerm->pColl;
+ if( pTerm->flags & EP_Collate ){
+ pColl = sqlite3ExprCollSeq(pParse, pTerm);
}else{
pColl = multiSelectCollSeq(pParse, p, aPermute[i]);
- pTerm->flags |= EP_ExpCollate;
- pTerm->pColl = pColl;
}
pKeyMerge->aColl[i] = pColl;
pKeyMerge->aSortOrder[i] = pOrderBy->a[i].sortOrder;
@@ -2608,9 +2606,6 @@
assert( pEList!=0 && pExpr->iColumn<pEList->nExpr );
assert( pExpr->pLeft==0 && pExpr->pRight==0 );
pNew = sqlite3ExprDup(db, pEList->a[pExpr->iColumn].pExpr, 0);
- if( pNew && pExpr->pColl ){
- pNew->pColl = pExpr->pColl;
- }
sqlite3ExprDelete(db, pExpr);
pExpr = pNew;
}
diff --git a/src/sqliteInt.h b/src/sqliteInt.h
index b9b9e6c..fc975bd 100644
--- a/src/sqliteInt.h
+++ b/src/sqliteInt.h
@@ -1678,7 +1678,6 @@
ExprList *pList; /* Function arguments or in "<expr> IN (<expr-list)" */
Select *pSelect; /* Used for sub-selects and "<expr> IN (<select>)" */
} x;
- CollSeq *pColl; /* The collation type of the column or 0 */
/* If the EP_Reduced flag is set in the Expr.flags mask, then no
** space is allocated for the fields below this point. An attempt to
@@ -1714,7 +1713,7 @@
#define EP_VarSelect 0x0020 /* pSelect is correlated, not constant */
#define EP_DblQuoted 0x0040 /* token.z was originally in "..." */
#define EP_InfixFunc 0x0080 /* True for an infix function: LIKE, GLOB, etc */
-#define EP_ExpCollate 0x0100 /* Collating sequence specified explicitly */
+#define EP_Collate 0x0100 /* Tree contains a TK_COLLATE opeartor */
#define EP_FixedDest 0x0200 /* Result needed in a specific register */
#define EP_IntValue 0x0400 /* Integer value contained in u.iValue */
#define EP_xIsSelect 0x0800 /* x.pSelect is valid (otherwise x.pList is) */
@@ -3022,7 +3021,6 @@
CollSeq *sqlite3FindCollSeq(sqlite3*,u8 enc, const char*,int);
CollSeq *sqlite3LocateCollSeq(Parse *pParse, const char*zName);
CollSeq *sqlite3ExprCollSeq(Parse *pParse, Expr *pExpr);
-Expr *sqlite3ExprSetColl(Expr*, CollSeq*);
Expr *sqlite3ExprSetCollByToken(Parse *pParse, Expr*, Token*);
int sqlite3CheckCollSeq(Parse *, CollSeq *);
int sqlite3CheckObjectName(Parse *, const char *);
diff --git a/src/where.c b/src/where.c
index 196d8f5..02fb722 100644
--- a/src/where.c
+++ b/src/where.c
@@ -563,23 +563,32 @@
** Commute a comparison operator. Expressions of the form "X op Y"
** are converted into "Y op X".
**
-** If a collation sequence is associated with either the left or right
+** If left/right precendence rules come into play when determining the
+** collating
** side of the comparison, it remains associated with the same side after
** the commutation. So "Y collate NOCASE op X" becomes
-** "X collate NOCASE op Y". This is because any collation sequence on
+** "X op Y". This is because any collation sequence on
** the left hand side of a comparison overrides any collation sequence
-** attached to the right. For the same reason the EP_ExpCollate flag
+** attached to the right. For the same reason the EP_Collate flag
** is not commuted.
*/
static void exprCommute(Parse *pParse, Expr *pExpr){
- u16 expRight = (pExpr->pRight->flags & EP_ExpCollate);
- u16 expLeft = (pExpr->pLeft->flags & EP_ExpCollate);
+ u16 expRight = (pExpr->pRight->flags & EP_Collate);
+ u16 expLeft = (pExpr->pLeft->flags & EP_Collate);
assert( allowedOp(pExpr->op) && pExpr->op!=TK_IN );
- pExpr->pRight->pColl = sqlite3ExprCollSeq(pParse, pExpr->pRight);
- pExpr->pLeft->pColl = sqlite3ExprCollSeq(pParse, pExpr->pLeft);
- SWAP(CollSeq*,pExpr->pRight->pColl,pExpr->pLeft->pColl);
- pExpr->pRight->flags = (pExpr->pRight->flags & ~EP_ExpCollate) | expLeft;
- pExpr->pLeft->flags = (pExpr->pLeft->flags & ~EP_ExpCollate) | expRight;
+ if( expRight==expLeft ){
+ /* Either X and Y both have COLLATE operator or neither do */
+ if( expRight ){
+ /* Both X and Y have COLLATE operators. Make sure X is always
+ ** used by clearing the EP_Collate flag from Y. */
+ pExpr->pRight->flags &= ~EP_Collate;
+ }else if( sqlite3ExprCollSeq(pParse, pExpr->pLeft)!=0 ){
+ /* Neither X nor Y have COLLATE operators, but X has a non-default
+ ** collating sequence. So add the EP_Collate marker on X to cause
+ ** it to be searched first. */
+ pExpr->pLeft->flags |= EP_Collate;
+ }
+ }
SWAP(Expr*,pExpr->pRight,pExpr->pLeft);
if( pExpr->op>=TK_GT ){
assert( TK_LT==TK_GT+2 );
@@ -1314,7 +1323,7 @@
Expr *pNewExpr2;
int idxNew1;
int idxNew2;
- CollSeq *pColl; /* Collating sequence to use */
+ Token sCollSeqName; /* Name of collating sequence */
pLeft = pExpr->x.pList->a[1].pExpr;
pStr2 = sqlite3ExprDup(db, pStr1, 0);
@@ -1336,16 +1345,19 @@
}
*pC = c + 1;
}
- pColl = sqlite3FindCollSeq(db, SQLITE_UTF8, noCase ? "NOCASE" : "BINARY",0);
+ sCollSeqName.z = noCase ? "NOCASE" : "BINARY";
+ sCollSeqName.n = 6;
+ pNewExpr1 = sqlite3ExprDup(db, pLeft, 0);
pNewExpr1 = sqlite3PExpr(pParse, TK_GE,
- sqlite3ExprSetColl(sqlite3ExprDup(db,pLeft,0), pColl),
- pStr1, 0);
+ sqlite3ExprSetCollByToken(pParse,pNewExpr1,&sCollSeqName),
+ pStr1, 0);
idxNew1 = whereClauseInsert(pWC, pNewExpr1, TERM_VIRTUAL|TERM_DYNAMIC);
testcase( idxNew1==0 );
exprAnalyze(pSrc, pWC, idxNew1);
+ pNewExpr2 = sqlite3ExprDup(db, pLeft, 0);
pNewExpr2 = sqlite3PExpr(pParse, TK_LT,
- sqlite3ExprSetColl(sqlite3ExprDup(db,pLeft,0), pColl),
- pStr2, 0);
+ sqlite3ExprSetCollByToken(pParse,pNewExpr2,&sCollSeqName),
+ pStr2, 0);
idxNew2 = whereClauseInsert(pWC, pNewExpr2, TERM_VIRTUAL|TERM_DYNAMIC);
testcase( idxNew2==0 );
exprAnalyze(pSrc, pWC, idxNew2);