The RIGHT JOIN logic added to version 3.39.0 added a new restriction on the

query flattener which resulted in a performance regression for some queries, as documented by [forum:/forumpost/96b9e5709cf47cda|forum post 96b9e5709cf47cda]. This change removes that restriction (thus restoring performance) and fixes the problem that RIGHT JOIN was having in a different way. FossilOrigin-Name: 501609eddf2a46d51e8cd56477a22d4df142a166e78fe4322c0db11d1eed6687
2026-01-06 08:01:16 +03:00 · 2022-09-21 09:54:41 +00:00
parent 703c208d30 faaacd3fbb
commit 2897b88f9e
7 changed files with 141 additions and 88 deletions
--- a/21
+++ b/21
@@ -1,5 +1,5 @@
-C Update\sthe\sfuzz\sinvariant\schecker\sto\sconform\sto\sthe\slatest\schanges\sin\ndbsqlfuzz.
-D 2022-09-20T17:21:54.866
+C The\sRIGHT\sJOIN\slogic\sadded\sto\sversion\s3.39.0\sadded\sa\snew\srestriction\son\sthe\nquery\sflattener\swhich\sresulted\sin\sa\sperformance\sregression\sfor\ssome\squeries,\nas\sdocumented\sby\s[forum:/forumpost/96b9e5709cf47cda|forum\spost\s96b9e5709cf47cda].\nThis\schange\sremoves\sthat\srestriction\s(thus\srestoring\sperformance)\sand\sfixes\sthe\nproblem\sthat\sRIGHT\sJOIN\swas\shaving\sin\sa\sdifferent\sway.
+D 2022-09-21T09:54:41.195
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
 F LICENSE.md df5091916dbb40e6e9686186587125e1b2ff51f022cc334e886c19a0e9982724
@@ -539,7 +539,7 @@ F src/date.c 94ce83b4cd848a387680a5f920c9018c16655db778c4d36525af0a0f34679ac5
 F src/dbpage.c 5808e91bc27fa3981b028000f8fadfdc10ce9e59a34ce7dc4e035a69be3906ec
 F src/dbstat.c 861e08690fcb0f2ee1165eff0060ea8d4f3e2ea10f80dab7d32ad70443a6ff2d
 F src/delete.c 86573edae75e3d3e9a8b590d87db8e47222103029df4f3e11fa56044459b514e
-F src/expr.c 24e828db6b2fab8aabfb5d2c0d83dbdfc5a1972b1147fa893350e317ab7e282f
+F src/expr.c 1cbdd76eeedb729ea9060df03e3e6b74a302784a13bfa38794a8194f894641ea
 F src/fault.c 460f3e55994363812d9d60844b2a6de88826e007
 F src/fkey.c 722f20779f5342a787922deded3628d8c74b5249cab04098cf17ee2f2aaff002
 F src/func.c 8f72e88cccdee22185133c10f96ccd61dc34c5ea4b1fa9a73c237ef59b2e64f1
@@ -588,7 +588,7 @@ F src/printf.c e99ee9741e79ae3873458146f59644276657340385ade4e76a5f5d1c25793764
 F src/random.c 546d6feb15ec69c1aafe9bb351a277cbb498fd5410e646add673acb805714960
 F src/resolve.c efea4e5fbecfd6d0a9071b0be0d952620991673391b6ffaaf4c277b0bb674633
 F src/rowset.c ba9515a922af32abe1f7d39406b9d35730ed65efab9443dc5702693b60854c92
-F src/select.c dcfd2df77e4601555cef48174e3c61723f72fccfdd5b80caa5da63e23338c040
+F src/select.c bb18acf4eded647fef88d4d543c673874dbebff516fbeba90a85e6c13f2a58cd
 F src/shell.c.in e7e7c2c69ae86c5ee9e8ad66227203d46ff6dce8700a1b1dababff01c71d33df
 F src/sqlite.h.in b9b7fd73239d94db20332bb6e504688001e5564b655e1318a4427a1caef4b99e
 F src/sqlite3.rc 5121c9e10c3964d5755191c80dd1180c122fc3a8
@@ -660,7 +660,7 @@ F src/upsert.c 8789047a8f0a601ea42fa0256d1ba3190c13746b6ba940fe2d25643a7e991937
 F src/utf.c ee39565f0843775cc2c81135751ddd93eceb91a673ea2c57f61c76f288b041a0
 F src/util.c 0be191521ff6d2805995f4910f0b6231b42843678b2efdc1abecaf39929a673f
 F src/vacuum.c bb346170b0b54c6683bba4a5983aea40485597fdf605c87ec8bc2e199fe88cd8
-F src/vdbe.c d27ec9a57f752fc2acf6a64d43bbf6072d2415efc976184f6d8a146e65819d3b
+F src/vdbe.c 0d1e3c658d98a7bb7201532ea7a3e4d59bf9165421c780d5f84c361e372f1179
 F src/vdbe.h 64619af62603dc3c4f5ff6ff6d2c8f389abd667a29ce6007ed44bd22b3211cd0
 F src/vdbeInt.h 17b7461ffcf9ee760d1341731715a419f6b8c763089a7ece25c2e8098d702b3f
 F src/vdbeapi.c fc3183daf72808b4311b228989120fdbc2dc44972fb0d77d5c453460cc0e5b2c
@@ -675,7 +675,7 @@ F src/vxworks.h d2988f4e5a61a4dfe82c6524dd3d6e4f2ce3cdb9
 F src/wal.c b9df133a705093da8977da5eb202eaadb844839f1c7297c08d33471f5491843d
 F src/wal.h c3aa7825bfa2fe0d85bef2db94655f99870a285778baa36307c0a16da32b226a
 F src/walker.c f890a3298418d7cba3b69b8803594fdc484ea241206a8dfa99db6dd36f8cbb3b
-F src/where.c 424c42590b71968a9b81cd890df2671902028613fee38a50ed4c2f7ca65315d3
+F src/where.c 3774d9831a6f4ac81f201fa9e2ec1bb40c99b7207dce7663c2e39fb73cc1bfd5
 F src/whereInt.h 70cd30de9ed784aa33fa6bd1245f060617de7a00d992469b6d8e419eed915743
 F src/wherecode.c 6bb1cf9d0a4e3e04dab0bf0ea4a8d936a0dcc05a7e2207beeda6c61aea6dd341
 F src/whereexpr.c 55a39f42aaf982574fbf52906371a84cceed98a994422198dfd03db4fce4cc46
@@ -1180,7 +1180,7 @@ F test/ioerr4.test f130fe9e71008577b342b8874d52984bd04ede2c
 F test/ioerr5.test 2edfa4fb0f896f733071303b42224df8bedd9da4
 F test/ioerr6.test a395a6ab144b26a9e3e21059a1ab6a7149cca65b
 F test/istrue.test e7f285bb70282625c258e866ce6337d4c762922f5a300e1b50f958aef6e7d9c9
-F test/join.test 21dbc65ab2476b10ae3ed1dcf64a99fb9a40473caa22f4cec2d6a829933c1442
+F test/join.test e32cb9b1491eed682489e2cde33a22a4eb7611fe5aa3b0aa4b275fe27ab3f3ac
 F test/join2.test 466b07233820f5deee66a6c3bf6e4500c8bbf7b83649e67606f5f649c07928c0
 F test/join3.test 6f0c774ff1ba0489e6c88a3e77b9d3528fb4fda0
 F test/join4.test 1a352e4e267114444c29266ce79e941af5885916
@@ -2000,8 +2000,9 @@ F vsixtest/vsixtest.tcl 6a9a6ab600c25a91a7acc6293828957a386a8a93
 F vsixtest/vsixtest.vcxproj.data 2ed517e100c66dc455b492e1a33350c1b20fbcdc
 F vsixtest/vsixtest.vcxproj.filters 37e51ffedcdb064aad6ff33b6148725226cd608e
 F vsixtest/vsixtest_TemporaryKey.pfx e5b1b036facdb453873e7084e1cae9102ccc67a0
-P 2647d73bb1876d01851f52cb3fd6af08a962ab336b6d51c160d27b1baf94f75f
-R 4c12344a007c9e9bdcdbf323756fa10a
+P 8d8c124349d2a25200ace9a60d76751594b722468ad0c39d927a5d2c6e2464e0 c7fbc9b0453050e2746af27e3a11e0c3701bef8c56da8e19173242c6ea3aff8b
+R ffb250842cc4325c51d9fb2f2a7cd199
+T +closed c7fbc9b0453050e2746af27e3a11e0c3701bef8c56da8e19173242c6ea3aff8b
 U drh
-Z 7b99665ad0fc78591cd70e35a24e34f9
+Z c4277c5174933f302dd59d96a8e6bf4e
 # Remove this line to create a well-formed Fossil manifest.
--- a/manifest.uuid
+++ b/manifest.uuid
@@ -1 +1 @@
-8d8c124349d2a25200ace9a60d76751594b722468ad0c39d927a5d2c6e2464e0
+501609eddf2a46d51e8cd56477a22d4df142a166e78fe4322c0db11d1eed6687
--- a/src/expr.c
+++ b/src/expr.c
@@ -3223,6 +3223,7 @@ void sqlite3CodeRhsOfIN(
    sqlite3VdbeChangeP4(v, addr, (void *)pKeyInfo, P4_KEYINFO);
  }
  if( addrOnce ){
+    sqlite3VdbeAddOp1(v, OP_NullRow, iTab);
    sqlite3VdbeJumpHere(v, addrOnce);
    /* Subroutine return */
    assert( ExprUseYSub(pExpr) );
--- a/src/select.c
+++ b/src/select.c
@@ -3724,7 +3724,7 @@ static int multiSelectOrderBy(
 ** the left operands of a RIGHT JOIN.  In either case, we need to potentially
 ** bypass the substituted expression with OP_IfNullRow.
 **
-** Suppose the original expression integer constant.  Even though the table
+** Suppose the original expression is an integer constant. Even though the table
 ** has the nullRow flag set, because the expression is an integer constant,
 ** it will not be NULLed out.  So instead, we insert an OP_IfNullRow opcode
 ** that checks to see if the nullRow flag is set on the table.  If the nullRow
@@ -4188,11 +4188,6 @@ static void renumberCursors(
 **
 **  (28)  The subquery is not a MATERIALIZED CTE.
 **
-**  (29)  Either the subquery is not the right-hand operand of a join with an
-**        ON or USING clause nor the right-hand operand of a NATURAL JOIN, or
-**        the right-most table within the FROM clause of the subquery
-**        is not part of an outer join.
-**
 **
 ** In this routine, the "p" parameter is a pointer to the outer query.
 ** The subquery is p->pSrc->a[iFrom].  isAgg is true if the outer query
@@ -4296,15 +4291,6 @@ static int flattenSubquery(
    }
    isOuterJoin = 1;
  }
-#ifdef SQLITE_EXTRA_IFNULLROW
-  else if( iFrom>0 && !isAgg ){
-    /* Setting isOuterJoin to -1 causes OP_IfNullRow opcodes to be generated for
-    ** every reference to any result column from subquery in a join, even
-    ** though they are not necessary.  This will stress-test the OP_IfNullRow 
-    ** opcode. */
-    isOuterJoin = -1;
-  }
-#endif

  assert( pSubSrc->nSrc>0 );  /* True by restriction (7) */
  if( iFrom>0 && (pSubSrc->a[0].fg.jointype & JT_LTORJ)!=0 ){
@@ -4314,33 +4300,6 @@ static int flattenSubquery(
    return 0;       /* (28) */
  }

-  /* Restriction (29): 
-  **
-  ** We do not want two constraints on the same FROM-clause term of the
-  ** flattened query where one constraint has the EP_InnerON flag and the
-  ** other has the EP_OuterON flag.
-  **
-  ** To prevent this, one or the other of the following conditions must be
-  ** false:
-  **
-  **   (29a)  The right-most entry in the FROM clause of the subquery
-  **          must not be part of an outer join.
-  **
-  **   (29b)  The subquery itself must not be the right operand of a 
-  **          NATURAL join or a join that has an ON or USING clause.
-  */
-  if( pSubSrc->nSrc>=2
-   && (pSubSrc->a[pSubSrc->nSrc-1].fg.jointype & JT_OUTER)!=0
-  ){
-    if( (pSubitem->fg.jointype & JT_NATURAL)!=0
-     || pSubitem->fg.isUsing
-     || NEVER(pSubitem->u3.pOn!=0) /* ON clause already shifted into WHERE */
-     || pSubitem->fg.isOn
-    ){
-      return 0;
-    }
-  }
-
  /* Restriction (17): If the sub-query is a compound SELECT, then it must
  ** use only the UNION ALL operator. And none of the simple select queries
  ** that make up the compound SELECT are allowed to be aggregate or distinct
--- a/src/vdbe.c
+++ b/src/vdbe.c
@@ -4897,12 +4897,16 @@ case OP_SeekHit: {
 /* Opcode: IfNotOpen P1 P2 * * *
 ** Synopsis: if( !csr[P1] ) goto P2
 **
-** If cursor P1 is not open, jump to instruction P2. Otherwise, fall through.
+** If cursor P1 is not open or if P1 is set to a NULL row using the
+** OP_NullRow opcode, then jump to instruction P2. Otherwise, fall through.
 */
 case OP_IfNotOpen: {        /* jump */
+  VdbeCursor *pCur;
+
  assert( pOp->p1>=0 && pOp->p1<p->nCursor );
-  VdbeBranchTaken(p->apCsr[pOp->p1]==0, 2);
-  if( !p->apCsr[pOp->p1] ){
+  pCur = p->apCsr[pOp->p1];
+  VdbeBranchTaken(pCur==0 || pCur->nullRow, 2);
+  if( pCur==0 || pCur->nullRow ){
    goto jump_to_p2_and_check_for_interrupt;
  }
  break;
--- a/src/where.c
+++ b/src/where.c
@@ -743,6 +743,38 @@ static void whereTraceIndexInfoOutputs(sqlite3_index_info *p){
 #define whereTraceIndexInfoOutputs(A)
 #endif

+/*
+** We know that pSrc is an operand of an outer join.  Return true if
+** pTerm is a constraint that is compatible with that join.
+**
+** pTerm must be EP_OuterON if pSrc is the right operand of an
+** outer join.  pTerm can be either EP_OuterON or EP_InnerON if pSrc
+** is the left operand of a RIGHT join.
+*/
+static int constraintCompatibleWithOuterJoin(
+  const WhereTerm *pTerm,       /* WHERE clause term to check */
+  const SrcItem *pSrc           /* Table we are trying to access */
+){
+  assert( (pSrc->fg.jointype&(JT_LEFT|JT_LTORJ|JT_RIGHT))!=0 ); /* By caller */
+  testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_LEFT );
+  testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_LTORJ );
+  testcase( ExprHasProperty(pTerm->pExpr, EP_OuterON) )
+  testcase( ExprHasProperty(pTerm->pExpr, EP_InnerON) );
+  if( !ExprHasProperty(pTerm->pExpr, EP_OuterON|EP_InnerON)
+   || pTerm->pExpr->w.iJoin != pSrc->iCursor
+  ){
+    return 0;
+  }
+  if( (pSrc->fg.jointype & (JT_LEFT|JT_RIGHT))!=0
+   && ExprHasProperty(pTerm->pExpr, EP_InnerON)
+  ){
+    return 0;
+  }
+  return 1;
+}
+ 
+
+
 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
 /*
 ** Return TRUE if the WHERE clause term pTerm is of a form where it
@@ -758,16 +790,10 @@ static int termCanDriveIndex(
  if( pTerm->leftCursor!=pSrc->iCursor ) return 0;
  if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0;
  assert( (pSrc->fg.jointype & JT_RIGHT)==0 );
-  if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))!=0 ){
-    testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_LEFT );
-    testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_LTORJ );
-    testcase( ExprHasProperty(pTerm->pExpr, EP_OuterON) )
-    testcase( ExprHasProperty(pTerm->pExpr, EP_InnerON) );
-    if( !ExprHasProperty(pTerm->pExpr, EP_OuterON|EP_InnerON)
-     || pTerm->pExpr->w.iJoin != pSrc->iCursor
-    ){
-      return 0;  /* See tag-20191211-001 */
-    }
+  if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))!=0
+   && !constraintCompatibleWithOuterJoin(pTerm,pSrc)
+  ){
+    return 0;
  }
  if( (pTerm->prereqRight & notReady)!=0 ) return 0;
  assert( (pTerm->eOperator & (WO_OR|WO_AND))==0 );
@@ -1184,17 +1210,10 @@ static sqlite3_index_info *allocateIndexInfo(
    ** right-hand table of a LEFT JOIN nor to the either table of a
    ** RIGHT JOIN.  See tag-20191211-001 for the
    ** equivalent restriction for ordinary tables. */
-    if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))!=0 ){
-      testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_LEFT );
-      testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_RIGHT );
-      testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_LTORJ );
-      testcase( ExprHasProperty(pTerm->pExpr, EP_OuterON) );
-      testcase( ExprHasProperty(pTerm->pExpr, EP_InnerON) );
-      if( !ExprHasProperty(pTerm->pExpr, EP_OuterON|EP_InnerON)
-       || pTerm->pExpr->w.iJoin != pSrc->iCursor
-      ){
-        continue;
-      }
+    if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))!=0
+     && !constraintCompatibleWithOuterJoin(pTerm,pSrc)
+    ){
+      continue;
    }
    nTerm++;
    pTerm->wtFlags |= TERM_OK;
@@ -2865,17 +2884,10 @@ static int whereLoopAddBtreeIndex(
    ** 2022-06-10: The same condition applies to termCanDriveIndex() above.
    ** https://sqlite.org/forum/forumpost/51e6959f61
    */
-    if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))!=0 ){
-      testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_LEFT );
-      testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_RIGHT );
-      testcase( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))==JT_LTORJ );
-      testcase( ExprHasProperty(pTerm->pExpr, EP_OuterON) )
-      testcase( ExprHasProperty(pTerm->pExpr, EP_InnerON) );
-      if( !ExprHasProperty(pTerm->pExpr, EP_OuterON|EP_InnerON)
-       || pTerm->pExpr->w.iJoin != pSrc->iCursor
-      ){
-        continue;
-      }
+    if( (pSrc->fg.jointype & (JT_LEFT|JT_LTORJ|JT_RIGHT))!=0
+     && !constraintCompatibleWithOuterJoin(pTerm,pSrc)
+    ){
+      continue;
    }

    if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){
--- a/test/join.test
+++ b/test/join.test
@@ -1130,5 +1130,81 @@ do_execsql_test join-27.5 {
    ) AS t99 ON b IN (1,2,3);
 } {}

+db null NULL
+do_execsql_test join-27.6 {
+  INSERT INTO t1 VALUES(3,4,NULL);
+  INSERT INTO t2 VALUES(1,2);
+  WITH t99(b) AS (
+    SELECT coalesce(b,3) FROM t2 AS x LEFT JOIN t1 ON c IN (SELECT x FROM t3)
+  )
+  SELECT d, e, b FROM t2 JOIN t99 ON b IN (1,2,3) ORDER BY +d;
+} {NULL NULL 3 NULL NULL 3 1 2 3 1 2 3}
+do_execsql_test join-27.7 {
+  SELECT d, e, b2
+    FROM t2 
+         JOIN (SELECT coalesce(b,3) AS b2 FROM t2 AS x LEFT JOIN t1
+                ON c IN (SELECT x FROM t3)) AS t99
+            ON b2 IN (1,2,3) ORDER BY +d;
+} {NULL NULL 3 NULL NULL 3 1 2 3 1 2 3}
+
+do_execsql_test join-27.8 {
+  DELETE FROM t1;
+  DELETE FROM t2 WHERE d IS NOT NULL;
+  DELETE FROM t3;
+  SELECT * FROM t2 JOIN (SELECT b FROM t2 LEFT JOIN t1
+                       ON c IN (SELECT x FROM t3)) AS t99 ON b IN (1,2,3);
+} {}
+
+do_execsql_test join-27.9 {
+  DELETE FROM t1;
+  DELETE FROM t2;
+  DELETE FROM t3;
+  INSERT INTO t1 VALUES(4,3,5);
+  INSERT INTO t2 VALUES(1,2);
+  INSERT INTO t3 VALUES(5);
+  SELECT * FROM t2 JOIN (SELECT b FROM t2 LEFT JOIN t1
+                       ON c IN (SELECT x FROM t3)) AS t99 ON b IS NULL;
+} {}
+do_execsql_test join-27.10 {
+  WITH t99(b) AS (
+    SELECT b FROM t2 AS x LEFT JOIN t1 ON c IN (SELECT x FROM t3)
+  )
+  SELECT d, e, b FROM t2 JOIN t99 ON b IS NULL;
+} {}
+
+
+# 2022-09-19 https://sqlite.org/forum/forumpost/96b9e5709cf47cda
+# Performance regression relative to version 3.38.0 that resulted from
+# a new query flattener restriction that was added to fixes the join-27.*
+# tests above.  The restriction needed to be removed and the join-27.*
+# problem fixed another way.
+#
+reset_db
+do_execsql_test join-28.1 {
+  CREATE TABLE t1(a INTEGER PRIMARY KEY, b INT, c INT);
+  CREATE TABLE t2(d INTEGER PRIMARY KEY, e INT);
+  CREATE VIEW t3(a,b,c,d,e) AS SELECT * FROM t1 LEFT JOIN t2 ON d=c;
+  CREATE TABLE t4(x INT, y INT);
+  INSERT INTO t1 VALUES(1,2,3);
+  INSERT INTO t2 VALUES(1,5);
+  INSERT INTO t4 VALUES(1,4);
+  SELECT a, b, y FROM t4 JOIN t3 ON a=x;
+} {1 2 4}
+do_eqp_test join-28.2 {
+  SELECT a, b, y FROM t4 JOIN t3 ON a=x;
+} {
+  QUERY PLAN
+  |--SCAN t4
+  `--SEARCH t1 USING INTEGER PRIMARY KEY (rowid=?)
+}
+# ^^^^^^^ Without the fix (if the query flattening optimization does not
+# run) the query plan above would look like this:
+#
+#   QUERY PLAN
+#   |--MATERIALIZE t3
+#   |  |--SCAN t1
+#   |  `--SEARCH t2 USING INTEGER PRIMARY KEY (rowid=?) LEFT-JOIN
+#   |--SCAN t4
+#   `--SEARCH t3 USING AUTOMATIC COVERING INDEX (a=?)

 finish_test