Clarification of the operation of the OR-term optimizer in where.c. (CVS 6730)

FossilOrigin-Name: 6b42dc3d04e98f91c203c277926ed6ead62a9270

manifest

@@ -1,5 +1,5 @@
-C Change\sthe\sbtree\sbalance\scode\sso\sthat\sit\sdoes\snot\scall\sbalance_nonroot()\srecursively.\s(CVS\s6729)
-D 2009-06-08T14:49:46
+C Clarification\sof\sthe\soperation\sof\sthe\sOR-term\soptimizer\sin\swhere.c.\s(CVS\s6730)
+D 2009-06-08T17:11:08
 F Makefile.arm-wince-mingw32ce-gcc fcd5e9cd67fe88836360bb4f9ef4cb7f8e2fb5a0
 F Makefile.in 8b8fb7823264331210cddf103831816c286ba446
 F Makefile.linux-gcc d53183f4aa6a9192d249731c90dbdffbd2c68654
@@ -212,7 +212,7 @@ F src/vdbeblob.c c25d7e7bc6d5917feeb17270bd275fa771f26e5c
 F src/vdbemem.c 05183d46094aa99b8f8350e5761b9369dbef35a8
 F src/vtab.c e2f4c92df7d06330b151448718c4724742ff444b
 F src/walker.c ec4b9742a4077ef80346e2f9aaf0f44c2d95087a
-F src/where.c 01b8d4733f177f047014c0b391b4ecbcc867dfd8
+F src/where.c 21555aa8b48345f5a1c706cf8ca485012f2cea03
 F test/aggerror.test a867e273ef9e3d7919f03ef4f0e8c0d2767944f2
 F test/alias.test 4529fbc152f190268a15f9384a5651bbbabc9d87
 F test/all.test 14165b3e32715b700b5f0cbf8f6e3833dda0be45
@@ -733,7 +733,7 @@ F tool/speedtest2.tcl ee2149167303ba8e95af97873c575c3e0fab58ff
 F tool/speedtest8.c 2902c46588c40b55661e471d7a86e4dd71a18224
 F tool/speedtest8inst1.c 293327bc76823f473684d589a8160bde1f52c14e
 F tool/vdbe-compress.tcl 672f81d693a03f80f5ae60bfefacd8a349e76746
-P 456ea541d67221a573b63e264c83f4448f2ed569
-R 7ada04c9a78d367df09e91faed8b31d8
-U danielk1977
-Z 47cd44a7ec9eea431d4519541126ee12
+P 7863db904d6fc36417c923e3d135eb2c145b9013
+R 6ef9e53cd5ad2b3cbcbea9afea1bc9a7
+U drh
+Z aebb118da6ff464bb9146e5d93538e1c

manifest.uuid

@@ -1 +1 @@
-7863db904d6fc36417c923e3d135eb2c145b9013
+6b42dc3d04e98f91c203c277926ed6ead62a9270

src/where.c

@@ -16,7 +16,7 @@
 ** so is applicable.  Because this module is responsible for selecting
 ** indices, you might also think of this module as the "query optimizer".
 **
-** $Id: where.c,v 1.402 2009/06/07 23:45:11 drh Exp $
+** $Id: where.c,v 1.403 2009/06/08 17:11:08 drh Exp $
 */
 #include "sqliteInt.h"
 
@@ -892,6 +892,22 @@ static void exprAnalyzeOrTerm(
   ** chngToIN holds a set of tables that *might* satisfy case 1.  But
   ** we have to do some additional checking to see if case 1 really
   ** is satisfied.
+  **
+  ** chngToIN will hold either 0, 1, or 2 bits.  The 0-bit case means
+  ** that there is no possibility of transforming the OR clause into an
+  ** IN operator because one or more terms in the OR clause contain
+  ** something other than == on a column in the single table.  The 1-bit
+  ** case means that every term of the OR clause is of the form
+  ** "table.column=expr" for some single table.  The one bit that is set
+  ** will correspond to the common table.  We still need to check to make
+  ** sure the same column is used on all terms.  The 2-bit case is when
+  ** the all terms are of the form "table1.column=table2.column".  It
+  ** might be possible to form an IN operator with either table1.column
+  ** or table2.column as the LHS if either is common to every term of
+  ** the OR clause.
+  **
+  ** Note that terms of the form "table.column1=table.column2" (the
+  ** same table on both sizes of the ==) cannot be optimized.
   */
   if( chngToIN ){
     int okToChngToIN = 0;     /* True if the conversion to IN is valid */
@@ -910,18 +926,38 @@ static void exprAnalyzeOrTerm(
       for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){
         assert( pOrTerm->eOperator==WO_EQ );
         pOrTerm->wtFlags &= ~TERM_OR_OK;
-        if( pOrTerm->leftCursor==iColumn ) continue;
-        if( (chngToIN & getMask(pMaskSet, pOrTerm->leftCursor))==0 ) continue;
+        if( pOrTerm->leftCursor==iCursor ){
+          /* This is the 2-bit case and we are on the second iteration and
+          ** current term is from the first iteration.  So skip this term. */
+          assert( j==1 );
+          continue;
+        }
+        if( (chngToIN & getMask(pMaskSet, pOrTerm->leftCursor))==0 ){
+          /* This term must be of the form t1.a==t2.b where t2 is in the
+          ** chngToIN set but t1 is not.  This term will be either preceeded
+          ** or follwed by an inverted copy (t2.b==t1.a).  Skip this term 
+          ** and use its inversion. */
+          testcase( pOrTerm->wtFlags & TERM_COPIED );
+          testcase( pOrTerm->wtFlags & TERM_VIRTUAL );
+          assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) );
+          continue;
+        }
         iColumn = pOrTerm->u.leftColumn;
         iCursor = pOrTerm->leftCursor;
         break;
       }
       if( i<0 ){
+        /* No candidate table+column was found.  This can only occur
+        ** on the second iteration */
         assert( j==1 );
         assert( (chngToIN&(chngToIN-1))==0 );
-        assert( chngToIN==getMask(pMaskSet, iColumn) );
+        assert( chngToIN==getMask(pMaskSet, iCursor) );
         break;
       }
+      testcase( j==1 );
+
+      /* We have found a candidate table and column.  Check to see if that
+      ** table and column is common to every term in the OR clause */
       okToChngToIN = 1;
       for(; i>=0 && okToChngToIN; i--, pOrTerm++){
         assert( pOrTerm->eOperator==WO_EQ );