Fix an issue in Bloom filters on RHS subsqueries to IN operators.

See [forum:/forumpost/792a09cb3df9e69f|forum post 792a09cb3d] for
a description of the problem.  Also improve comments related
to [baa83b460c677c21] which was origin of the problem.

FossilOrigin-Name: cdef486e212fe4b26605065d9cff08f608cb80df48ee64e4be63637769bdfacc

1 files changed, 5 insertions, 4 deletions

diff --git a/src/expr.c b/src/expr.c
index dd5ea2038..bf15811bc 100644
--- a/src/expr.c
+++ b/src/expr.c
@@ -3626,11 +3626,12 @@ void sqlite3CodeRhsOfIN(
       sqlite3SelectDelete(pParse->db, pCopy);
       sqlite3DbFree(pParse->db, dest.zAffSdst);
       if( addrBloom ){
+        /* Remember that location of the Bloom filter in the P3 operand
+        ** of the OP_Once that began this subroutine. tag-202407032019 */
         sqlite3VdbeGetOp(v, addrOnce)->p3 = dest.iSDParm2;
         if( dest.iSDParm2==0 ){
-          sqlite3VdbeChangeToNoop(v, addrBloom);
-        }else{
-          sqlite3VdbeGetOp(v, addrOnce)->p3 = dest.iSDParm2;
+          /* If the Bloom filter won't actually be used, keep it small */
+          sqlite3VdbeGetOp(v, addrBloom)->p1 = 10;
         }
       }
       if( rc ){
@@ -4077,7 +4078,7 @@ static void sqlite3ExprCodeIN(
       if( ExprHasProperty(pExpr, EP_Subrtn) ){
         const VdbeOp *pOp = sqlite3VdbeGetOp(v, pExpr->y.sub.iAddr);
         assert( pOp->opcode==OP_Once || pParse->nErr );
-        if( pOp->opcode==OP_Once && pOp->p3>0 ){
+        if( pOp->opcode==OP_Once && pOp->p3>0 ){  /* tag-202407032019 */
           assert( OptimizationEnabled(pParse->db, SQLITE_BloomFilter) );
           sqlite3VdbeAddOp4Int(v, OP_Filter, pOp->p3, destIfFalse,
                                rLhs, nVector); VdbeCoverage(v);