Add the left join strength reduction optimization. Enhance the push-down

optimization so that it works with many LEFT JOINs. FossilOrigin-Name: dd568c27b1d7656388ea5b4132cc0265aedd7348d265d8e8c7412b00b28a31aa
2018-03-22 12:00:43 +00:00
parent 66651c2b4b 7fbb101c54
commit c6c5ab80ee
8 changed files with 168 additions and 58 deletions
@@ -1,5 +1,5 @@
-C Add\sthe\s--valid-sql\soption\sto\sthe\soptfuzz\stest\sprogram.
-D 2018-03-22T11:28:31.684
+C Add\sthe\sleft\sjoin\sstrength\sreduction\soptimization.\s\sEnhance\sthe\spush-down\noptimization\sso\sthat\sit\sworks\swith\smany\sLEFT\sJOINs.
+D 2018-03-22T12:00:43.103
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
 F Makefile.in 7016fc56c6b9bfe5daac4f34be8be38d8c0b5fab79ccbfb764d3b23bf1c6fff3
@@ -441,7 +441,7 @@ F src/date.c ebe1dc7c8a347117bb02570f1a931c62dd78f4a2b1b516f4837d45b7d6426957
 F src/dbpage.c 8db4c97f630e7d83f884ea75caf1ffd0988c160e9d530194d93721c80821e0f6
 F src/dbstat.c 7a4ba8518b6369ef3600c49cf9c918ad979acba610b2aebef1b656d649b96720
 F src/delete.c 20c8788451dc737a967c87ea53ad43544d617f5b57d32ccce8bd52a0daf9e89b
-F src/expr.c 904f85e5f5fa232488cbb662fcf3306ae203728751468670fb9840a9b1886abc
+F src/expr.c 21ba8e1f8b9bc035053941d4f8df3ca5376be5e63ff62f945c2bc4f9b8890362
 F src/fault.c 460f3e55994363812d9d60844b2a6de88826e007
 F src/fkey.c d617daf66b5515e2b42c1405b2b4984c30ca50fb705ab164271a9bf66c69e331
 F src/func.c 94f42cba2cc1c34aeaa441022ba0170ec3fec4bba54db4e0ded085c6dc0fdc51
@@ -489,12 +489,12 @@ F src/printf.c d3b7844ddeb11fbbdd38dd84d09c9c1ac171d21fb038473c3aa97981201cc660
 F src/random.c 80f5d666f23feb3e6665a6ce04c7197212a88384
 F src/resolve.c 66c73fcb7719b8ff0e841b58338f13604ff3e2b50a723f9b8f383595735262f6
 F src/rowset.c 7b7e7e479212e65b723bf40128c7b36dc5afdfac
-F src/select.c 69ed6ca9d1bf4adb41472b0740b7b90868438dedcdd5ab2fd608070eca30d073
+F src/select.c 2c408ab9b4bfc807e6a9d3d45337d20ebe72a8b7ac793854710bf9d572e7ab6d
 F src/shell.c.in 911b9e3bce40413c78fdba28efa28363e98183819bd4b300780bf57bacfc4b84
 F src/sqlite.h.in 19762b57baa1ade67531f254de94374428fb9c82452ef305017847945f9c2911
 F src/sqlite3.rc 5121c9e10c3964d5755191c80dd1180c122fc3a8
 F src/sqlite3ext.h 83a3c4ce93d650bedfd1aa558cb85a516bd6d094445ee989740827d0d944368d
-F src/sqliteInt.h 7e9deb145c110289f50fcf72a8742d030834885676482391fe83c3cee75f97d4
+F src/sqliteInt.h 541896393a2ac07de571bcb0ce05b9422b94a6fb3900b005fa095d70b4d9545a
 F src/sqliteLimit.h 1513bfb7b20378aa0041e7022d04acb73525de35b80b252f1b83fedb4de6a76b
 F src/status.c 46e7aec11f79dad50965a5ca5fa9de009f7d6bde08be2156f1538a0a296d4d0e
 F src/table.c b46ad567748f24a326d9de40e5b9659f96ffff34
@@ -736,7 +736,7 @@ F test/cse.test 277350a26264495e86b1785f34d2d0c8600e021c
 F test/csv01.test 526fc6aefd052badd5a0283f86b4b395c3df76bfe98d96c801f494f5e2c7836c
 F test/ctime.test 78749e6c9a5f0010d67985be80788f841e3cd2da18114e2ed6010399a7d807f3
 F test/cursorhint.test 7bc346788390475e77a345da2b92270d04d35856
-F test/cursorhint2.test 8457e93d97f665f23f97cdbc8477d16e3480331b
+F test/cursorhint2.test 0078ae1ded4afcf5eb80d06e3a72b6e1c3f1a646aab26eeb583b0a9ec6f0d56e
 F test/date.test 9b73bbeb1b82d9c1f44dec5cf563bf7da58d2373
 F test/date2.test 74c234bece1b016e94dd4ef9c8cc7a199a8806c0e2291cab7ba64bace6350b10
 F test/dbfuzz.c 73047c920d6210e5912c87cdffd9a1c281d4252e
@@ -771,7 +771,7 @@ F test/e_fts3.test 8cf40550bb088a6aa187c818c00fabe26ef82900a4cd5c66b427ccafe28be
 F test/e_insert.test f02f7f17852b2163732c6611d193f84fc67bc641fb4882c77a464076e5eba80e
 F test/e_reindex.test 2bebf7b393e519198b7c654407221cf171a439b8
 F test/e_resolve.test a61751c368b109db73df0f20fc75fb47e166b1d8
-F test/e_select.test 16651bb681e83a1a2875ff4a595ed2b4b4dee375
+F test/e_select.test 6fd45fd4a59ec82b6dda7468699dcc0ec1a72538577750b4f90357a62c1d2723
 F test/e_select2.test aceb80ab927d46fba5ce7586ebabf23e2bb0604f
 F test/e_totalchanges.test b12ee5809d3e63aeb83238dd501a7bca7fd72c10
 F test/e_update.test f46c2554d915c9197548681e8d8c33a267e84528
@@ -1004,7 +1004,7 @@ F test/ioerr5.test 2edfa4fb0f896f733071303b42224df8bedd9da4
 F test/ioerr6.test a395a6ab144b26a9e3e21059a1ab6a7149cca65b
 F test/istrue.test 2d89873d72248b7a0ef7b5457120100cce75a19181bcf4b30e90a3f96a5c5558
 F test/join.test 442c462eea85cf065d70a663c626b780a95af6e11585d909bb63b87598afe678
-F test/join2.test f1b5ba94ba70a791df2d99e61adcebf62edacad1eca57881c97eb3e2ca718595
+F test/join2.test 2339e3a8089178607a718c7fa1f6c3826579132396f5304a0f8e9dbefa60c84f
 F test/join3.test 6f0c774ff1ba0489e6c88a3e77b9d3528fb4fda0
 F test/join4.test 1a352e4e267114444c29266ce79e941af5885916
 F test/join5.test c6bd62effc37a152bea735f9ef241b19bb967bd4593dc99b20e2fc55ae707e38
@@ -1715,7 +1715,8 @@ F vsixtest/vsixtest.tcl 6a9a6ab600c25a91a7acc6293828957a386a8a93
 F vsixtest/vsixtest.vcxproj.data 2ed517e100c66dc455b492e1a33350c1b20fbcdc
 F vsixtest/vsixtest.vcxproj.filters 37e51ffedcdb064aad6ff33b6148725226cd608e
 F vsixtest/vsixtest_TemporaryKey.pfx e5b1b036facdb453873e7084e1cae9102ccc67a0
-P 21346bbce9fd161e8a2037834a6e8eb443a901109ddb3a52c1b7a29000ffeac8
-R 338666e76c5adcc84ca67c4306a8f3fb
+P a8dfeec73b069f2dd7371c2792b36f152945d32120843db5361ff7e04de177bc b5d3dd8cb0b1e42ed0671a12d22af05194ea9522e4f41fd4bb0deff70b8b0757
+R 829b4bcd2ce5d178bebfa1657a2d3cee
+T +closed b5d3dd8cb0b1e42ed0671a12d22af05194ea9522e4f41fd4bb0deff70b8b0757
 U drh
-Z 865cfc6e4e0c84a2365c19e9bd9ce995
+Z f7b572645a2b0a1f5c79c87cac8e08a9
@@ -1 +1 @@
-a8dfeec73b069f2dd7371c2792b36f152945d32120843db5361ff7e04de177bc
+dd568c27b1d7656388ea5b4132cc0265aedd7348d265d8e8c7412b00b28a31aa
@@ -5001,6 +5001,58 @@ int sqlite3ExprImpliesExpr(Parse *pParse, Expr *pE1, Expr *pE2, int iTab){
  return 0;
 }

+/*
+** This is the Expr node callback for sqlite3ExprImpliesNotNullRow().
+** If the expression node requires that the table at pWalker->iCur
+** have a non-NULL column, then set pWalker->eCode to 1 and abort.
+*/
+static int impliesNotNullRow(Walker *pWalker, Expr *pExpr){
+  if( ExprHasProperty(pExpr, EP_FromJoin) ) return WRC_Prune;
+  switch( pExpr->op ){
+    case TK_ISNULL:
+    case TK_IS:
+    case TK_OR:
+    case TK_FUNCTION:
+    case TK_AGG_FUNCTION:
+      return WRC_Prune;
+    case TK_COLUMN:
+    case TK_AGG_COLUMN:
+      if( pWalker->u.iCur==pExpr->iTable ){
+        pWalker->eCode = 1;
+        return WRC_Abort;
+      }
+      return WRC_Prune;
+    default:
+      return WRC_Continue;
+  }
+}
+
+/*
+** Return true (non-zero) if expression p can only be true if at least
+** one column of table iTab is non-null.  In other words, return true
+** if expression p will always be NULL or false if every column of iTab
+** is NULL.
+**
+** Terms of p that are marked with EP_FromJoin (and hence that come from
+** the ON or USING clauses of LEFT JOINS) are excluded from the analysis.
+**
+** This routine is used to check if a LEFT JOIN can be converted into
+** an ordinary JOIN.  The p argument is the WHERE clause.  If the WHERE
+** clause requires that some column of the right table of the LEFT JOIN
+** be non-NULL, then the LEFT JOIN can be safely converted into an
+** ordinary join.
+*/
+int sqlite3ExprImpliesNonNullRow(Expr *p, int iTab){
+  Walker w;
+  w.xExprCallback = impliesNotNullRow;
+  w.xSelectCallback = 0;
+  w.xSelectCallback2 = 0;
+  w.eCode = 0;
+  w.u.iCur = iTab;
+  sqlite3WalkExpr(&w, p);
+  return w.eCode;
+}
+
 /*
 ** An instance of the following structure is used by the tree walker
 ** to determine if an expression can be evaluated by reference to the
@@ -382,6 +382,29 @@ static void setJoinExpr(Expr *p, int iTable){
  } 
 }

+/* Undo the work of setJoinExpr().  In the expression tree p, convert every
+** term that is marked with EP_FromJoin and iRightJoinTable==iTable into
+** an ordinary term that omits the EP_FromJoin mark.
+**
+** This happens when a LEFT JOIN is simplified into an ordinary JOIN.
+*/
+static void unsetJoinExpr(Expr *p, int iTable){
+  while( p ){
+    if( ExprHasProperty(p, EP_FromJoin)
+     && (iTable<0 || p->iRightJoinTable==iTable) ){
+      ExprClearProperty(p, EP_FromJoin);
+    }
+    if( p->op==TK_FUNCTION && p->x.pList ){
+      int i;
+      for(i=0; i<p->x.pList->nExpr; i++){
+        unsetJoinExpr(p->x.pList->a[i].pExpr, iTable);
+      }
+    }
+    unsetJoinExpr(p->pLeft, iTable);
+    p = p->pRight;
+  } 
+}
+
 /*
 ** This routine processes the join information for a SELECT statement.
 ** ON and USING clauses are converted into extra terms of the WHERE clause.
@@ -3834,12 +3857,21 @@ static int flattenSubquery(
 **   (3) The inner query has a LIMIT clause (since the changes to the WHERE
 **       close would change the meaning of the LIMIT).
 **
-**   (4) The inner query is the right operand of a LEFT JOIN.  (The caller
-**       enforces this restriction since this routine does not have enough
-**       information to know.)
+**   (4) (** This restriction was removed on 2018-03-21.  It used to read:
+**       The inner query is the right operand of a LEFT JOIN. **)
 **
 **   (5) The WHERE clause expression originates in the ON or USING clause
-**       of a LEFT JOIN.
+**       of a LEFT JOIN where iCursor is not the right-hand table of that
+**       left join.  An example:
+**
+**           SELECT *
+**           FROM (SELECT 1 AS a1 UNION ALL SELECT 2) AS aa
+**           JOIN (SELECT 1 AS b2 UNION ALL SELECT 2) AS bb ON (a1=b2)
+**           LEFT JOIN (SELECT 8 AS c3 UNION ALL SELECT 9) AS cc ON (b2=2);
+**
+**       The correct answer is three rows:  (1,1,NULL),(2,2,8),(2,2,9).
+**       But if the (b2=2) term were to be pushed down into the bb subquery,
+**       then the (1,1,NULL) row would be suppressed.
 **
 ** Return 0 if no changes are made and non-zero if one or more WHERE clause
 ** terms are duplicated into the subquery.
@@ -3875,12 +3907,15 @@ static int pushDownWhereTerms(
    nChng += pushDownWhereTerms(pParse, pSubq, pWhere->pRight, iCursor);
    pWhere = pWhere->pLeft;
  }
-  if( ExprHasProperty(pWhere,EP_FromJoin) ) return 0; /* restriction (5) */
+  if( ExprHasProperty(pWhere,EP_FromJoin) && pWhere->iRightJoinTable!=iCursor ){
+    return 0; /* restriction (5) */
+  }
  if( sqlite3ExprIsTableConstant(pWhere, iCursor) ){
    nChng++;
    while( pSubq ){
      SubstContext x;
      pNew = sqlite3ExprDup(pParse->db, pWhere, 0);
+      unsetJoinExpr(pNew, -1);
      x.pParse = pParse;
      x.iTable = iCursor;
      x.iNewTable = iCursor;
@@ -5175,13 +5210,29 @@ int sqlite3Select(
    generateColumnNames(pParse, p);
  }

-  /* Try to flatten subqueries in the FROM clause up into the main query
+  /* Try to various optimizations (flattening subqueries, and strength
+  ** reduction of join operators) in the FROM clause up into the main query
  */
 #if !defined(SQLITE_OMIT_SUBQUERY) || !defined(SQLITE_OMIT_VIEW)
  for(i=0; !p->pPrior && i<pTabList->nSrc; i++){
    struct SrcList_item *pItem = &pTabList->a[i];
    Select *pSub = pItem->pSelect;
    Table *pTab = pItem->pTab;
+
+    /* Convert LEFT JOIN into JOIN if there are terms of the right table
+    ** of the LEFT JOIN used in the WHERE clause.
+    */
+    if( (pItem->fg.jointype & JT_LEFT)!=0
+     && sqlite3ExprImpliesNonNullRow(p->pWhere, pItem->iCursor)
+     && OptimizationEnabled(db, SQLITE_SimplifyJoin)
+    ){
+      SELECTTRACE(0x100,pParse,p,
+                ("LEFT-JOIN simplifies to JOIN on term %d\n",i));
+      pItem->fg.jointype &= ~(JT_LEFT|JT_OUTER);
+      unsetJoinExpr(p->pWhere, pItem->iCursor);
+    }
+
+    /* No futher action if this term of the FROM clause is no a subquery */
    if( pSub==0 ) continue;

    /* Catch mismatch in the declared columns of a view and the number of
@@ -5322,8 +5373,7 @@ int sqlite3Select(
    /* Make copies of constant WHERE-clause terms in the outer query down
    ** inside the subquery.  This can help the subquery to run more efficiently.
    */
-    if( (pItem->fg.jointype & JT_OUTER)==0
-     && OptimizationEnabled(db, SQLITE_PushDown)
+    if( OptimizationEnabled(db, SQLITE_PushDown)
     && pushDownWhereTerms(pParse, pSub, p->pWhere, pItem->iCursor)
    ){
 #if SELECTTRACE_ENABLED
@@ -1533,6 +1533,7 @@ struct sqlite3 {
 #define SQLITE_Stat34         0x0800   /* Use STAT3 or STAT4 data */
   /* TH3 expects the Stat34  ^^^^^^ value to be 0x0800.  Don't change it */
 #define SQLITE_PushDown       0x1000   /* The push-down optimization */
+#define SQLITE_SimplifyJoin   0x2000   /* Convert LEFT JOIN to JOIN */
 #define SQLITE_AllOpts        0xffff   /* All optimizations */

 /*
@@ -3823,6 +3824,7 @@ int sqlite3ExprCompare(Parse*,Expr*, Expr*, int);
 int sqlite3ExprCompareSkip(Expr*, Expr*, int);
 int sqlite3ExprListCompare(ExprList*, ExprList*, int);
 int sqlite3ExprImpliesExpr(Parse*,Expr*, Expr*, int);
+int sqlite3ExprImpliesNonNullRow(Expr*,int);
 void sqlite3ExprAnalyzeAggregates(NameContext*, Expr*);
 void sqlite3ExprAnalyzeAggList(NameContext*,ExprList*);
 int sqlite3ExprCoveredByIndex(Expr*, int iCur, Index *pIdx);
@@ -136,45 +136,50 @@ do_extract_hints_test 2.5 {
  x2 {EQ(c0,r[2])}
 }

-do_extract_hints_test 2.6 {
-  SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE 0 = (b IS NOT NULL)
-} {
-  x2 {EQ(c0,r[2])}
-}
-
-do_extract_hints_test 2.7 {
-  SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE 0 = (b IS NOT +NULL)
-} {
-  x2 {EQ(c0,r[2])}
-}
-
-do_extract_hints_test 2.8 {
-  SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE b IS NOT +NULL
-} {
-  x2 {EQ(c0,r[2])}
-}
-
-do_extract_hints_test 2.9 {
-  SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE CASE b WHEN 0 THEN 0 ELSE 1 END;
-} {
-  x2 {EQ(c0,r[2])}
-}
-
-do_extract_hints_test 2.10 {
-  SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE x2.b = 32+32
-} {
-  x2 {AND(EQ(c1,ADD(32,32)),EQ(c0,r[2]))}
-}
-
-ifcapable !icu {
-  # This test only works using the built-in LIKE, not the ICU LIKE extension.
-  do_extract_hints_test 2.11 {
-    SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE x2.b LIKE 'abc%'
+if {0} {
+  # These tests no longer work due to the LEFT-JOIN strength reduction
+  # optimization
+  do_extract_hints_test 2.6 {
+    SELECT * FROM x1 CROSS JOIN x2 ON (a=x) WHERE 0 = (b IS NOT NULL)
  } {
-    x2 {AND(expr,EQ(c0,r[2]))}
+    x2 {EQ(c0,r[2])}
+  }
+  
+  do_extract_hints_test 2.7 {
+    SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE 0 = (b IS NOT +NULL)
+  } {
+    x2 {EQ(c0,r[2])}
+  }
+  
+  do_extract_hints_test 2.8 {
+    SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE b IS NOT +NULL
+  } {
+    x2 {EQ(c0,r[2])}
+  }
+  
+  do_extract_hints_test 2.9 {
+    SELECT * FROM x1 LEFT JOIN x2 ON (a=x)
+      WHERE CASE b WHEN 0 THEN 0 ELSE 1 END;
+  } {
+    x2 {EQ(c0,r[2])}
+  }
+  
+  do_extract_hints_test 2.10 {
+    SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE x2.b = 32+32
+  } {
+    x2 {AND(EQ(c1,ADD(32,32)),EQ(c0,r[2]))}
+  }
+  
+  ifcapable !icu {
+    # This test only works using the built-in LIKE, not the ICU LIKE extension.
+    do_extract_hints_test 2.11 {
+      SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE x2.b LIKE 'abc%'
+    } {
+      x2 {AND(expr,EQ(c0,r[2]))}
+    }
  }
 }
-
+  
 do_extract_hints_test 2.12 {
  SELECT * FROM x1 LEFT JOIN x2 ON (a=x) WHERE coalesce(x2.b, 1)
 } {
@@ -748,7 +748,7 @@ do_execsql_test e_select-3.2.1a {
  SELECT k FROM x1 LEFT JOIN x2 USING(k)
 } {1 2 3 4 5 6}
 do_execsql_test e_select-3.2.1b {
-  SELECT k FROM x1 LEFT JOIN x2 USING(k) WHERE x2.k
+  SELECT k FROM x1 LEFT JOIN x2 USING(k) WHERE x2.k ORDER BY +k
 } {1 3 5}
 do_execsql_test e_select-3.2.2 {
  SELECT k FROM x1 LEFT JOIN x2 USING(k) WHERE x2.k IS NULL
@@ -86,7 +86,7 @@ do_execsql_test 2.0 {
 }

 do_catchsql_test 2.1 {
-  SELECT * FROM aa LEFT JOIN cc ON (a=b) JOIN bb ON (b=c);
+  SELECT * FROM aa LEFT JOIN cc ON (a=b) JOIN bb ON (b=coalesce(c,1));
 } {1 {ON clause references tables to its right}}
 do_catchsql_test 2.2 {
  SELECT * FROM aa JOIN cc ON (a=b) JOIN bb ON (b=c);