Materialize any CTE that is used more than once.

FossilOrigin-Name: ba59159fbe6b83fb6d79fbfee22d983768b0ebbaac7e99d2ac66c810e5e04100

manifest

@@ -1,5 +1,5 @@
-C Rename\sthe\s"struct\sSrcList_item"\sobject\sto\sthe\smore\ssuccinct\s"SrcItem".\nThis\sis\sa\ssymbolic\schange\sonly.\s\sThe\slogic\sis\sunmodified.
-D 2021-02-21T21:04:54.414
+C Materialize\sany\sCTE\sthat\sis\sused\smore\sthan\sonce.
+D 2021-02-21T23:44:14.598
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
 F LICENSE.md df5091916dbb40e6e9686186587125e1b2ff51f022cc334e886c19a0e9982724
@@ -493,8 +493,8 @@ F src/ctime.c 2a322b9a3d75771fb4d99e0702851f4f68dda982507a0f798eefb0712969a410
 F src/date.c dace306a10d9b02ee553d454c8e1cf8d3c9b932e137738a6b15b90253a9bfc10
 F src/dbpage.c 8a01e865bf8bc6d7b1844b4314443a6436c07c3efe1d488ed89e81719047833a
 F src/dbstat.c 3aa79fc3aed7ce906e4ea6c10e85d657299e304f6049861fe300053ac57de36c
-F src/delete.c 8097c2c8feade9c46a1828dbdc98e1f52f4934fbff38233a2e7b98729528d244
-F src/expr.c 3a756aae0435178ddaed3b4e99ed7c6006f94d380881a249b7cf9d3fca8a405e
+F src/delete.c 83dbb8d8c136a100b7a15da38c374b88d82460b3b14f036c169c2d5810223112
+F src/expr.c 6793c836aff149b14011ad546ae1648a18573779ee78f5a7d375f2a3047e8c8e
 F src/fault.c 460f3e55994363812d9d60844b2a6de88826e007
 F src/fkey.c e9063648396c58778f77583a678342fe4a9bc82436bf23c5f9f444f2df0fdaa4
 F src/func.c 479f6929be027eb0210cbdde9d3529c012facf082d64a6b854a9415940761e5e
@@ -537,17 +537,17 @@ F src/pcache.h 4f87acd914cef5016fae3030343540d75f5b85a1877eed1a2a19b9f284248586
 F src/pcache1.c 6596e10baf3d8f84cc1585d226cf1ab26564a5f5caf85a15757a281ff977d51a
 F src/pragma.c 6daaaecc26a4b09481d21722525b079ce756751a43a79cc1d8f122d686806193
 F src/pragma.h 8dc78ab7e9ec6ce3ded8332810a2066f1ef6267e2e03cd7356ee00276125c6cf
-F src/prepare.c 7a534d100c556f45f10aee131f2e4244cb52547b7cf17e1c393f55d8abb62e97
+F src/prepare.c f634a9e799a6b1c136d8ee12479cffa22862bfb807d307b1db406aa0cdb042a5
 F src/printf.c 10e61ec79dd9d41fdc77afee4e0df04fbb427f309c043118fe0b26a7d7db488a
 F src/random.c 80f5d666f23feb3e6665a6ce04c7197212a88384
 F src/resolve.c c263fa5b255a03314c2418f936386e903d01c3e7cbec25a363a586ef3f10b249
 F src/rowset.c ba9515a922af32abe1f7d39406b9d35730ed65efab9443dc5702693b60854c92
-F src/select.c b3277dc42ab20367b023c78d998119915403aa98c4e7230567dd9884ca38d7fd
+F src/select.c cc8df79d5838f7c1d0bdbcf3e78b3d2efb781a09a39c4e5146fe7a669a621e82
 F src/shell.c.in 844417f84df1f6c4fce1c815629a888cfdcf219e86513e9c332bbcc38832f477
 F src/sqlite.h.in 8855a19f37ade8dad189a9e48233a2ebe1b46faf469c7eb0906a654e252dcc57
 F src/sqlite3.rc 5121c9e10c3964d5755191c80dd1180c122fc3a8
 F src/sqlite3ext.h 61b38c073d5e1e96a3d45271b257aef27d0d13da2bea5347692ae579475cd95e
-F src/sqliteInt.h dbe0a5541c6eaabd3535ca6fd9838452bef7d25430267ac6701ef4d0c2cce46a
+F src/sqliteInt.h 4e6c2c06e0234ef5e3049e1d0fc551c45db202bee9d90e2397862ec1ae9ca2cf
 F src/sqliteLimit.h d7323ffea5208c6af2734574bae933ca8ed2ab728083caa117c9738581a31657
 F src/status.c 4b8bc2a6905163a38b739854a35b826c737333fab5b1f8e03fa7eb9a4799c4c1
 F src/table.c 0f141b58a16de7e2fbe81c308379e7279f4c6b50eb08efeec5892794a0ba30d1
@@ -607,7 +607,7 @@ F src/test_window.c cdae419fdcea5bad6dcd9368c685abdad6deb59e9fc8b84b153de513d394
 F src/test_wsd.c 41cadfd9d97fe8e3e4e44f61a4a8ccd6f7ca8fe9
 F src/threads.c 4ae07fa022a3dc7c5beb373cf744a85d3c5c6c3c
 F src/tokenize.c c64c49d7c2ec4490c2fef1f24350167ba16b03b0c6cee58ad1a1d70a4325d4e9
-F src/treeview.c 04bf47d19b9f6f0b636cf42fec6e676b140d781bae792f52a8a014867e838e3e
+F src/treeview.c c6260e1fa5f41c361b2409edc9b0050bcaef5bc4d6abc467fbc45f0d7ccf3d84
 F src/trigger.c 861c3ec2c5b0fc830bdf82470454a9324fad70cbaa96d2e208fb54577c9e8d28
 F src/update.c 0f5a61f0787199983530a33f6fffe4f52742f35fcdf6ccfad1078b1a8bc17723
 F src/upsert.c df8f1727d62b5987c4fd302cd4d7c0c84ae57cd65683c5a34a740dfe24039235
@@ -629,7 +629,7 @@ F src/vxworks.h d2988f4e5a61a4dfe82c6524dd3d6e4f2ce3cdb9
 F src/wal.c 69e770e96fd56cc21608992bf2c6f1f3dc5cf2572d0495c6a643b06c3a679f14
 F src/wal.h c3aa7825bfa2fe0d85bef2db94655f99870a285778baa36307c0a16da32b226a
 F src/walker.c d42d6c80ea363ef689a462e65eefcfe87deab924c50de5baa37ecb6af7d7ddaa
-F src/where.c 40735d2b96b9b071a510fd3f090a195f48a09beb0c97ae0a11fd49e7c9396af9
+F src/where.c a02138440d7230493b5e508664d629f3e1e7615737a5d83aac3a2955d3a654ff
 F src/whereInt.h 446e5e8018f83358ef917cf32d8e6a86dc8430113d0b17e720f1839d3faa44c4
 F src/wherecode.c e57a8690311a75d06e723e8d379f9831de04aba300e07174d236e32a7f9c7a13
 F src/whereexpr.c 2dc51263e1fb8d8723e97a077a9a137ab0534e59e4cb88b3195f65edbe43cc32
@@ -1905,7 +1905,10 @@ F vsixtest/vsixtest.tcl 6a9a6ab600c25a91a7acc6293828957a386a8a93
 F vsixtest/vsixtest.vcxproj.data 2ed517e100c66dc455b492e1a33350c1b20fbcdc
 F vsixtest/vsixtest.vcxproj.filters 37e51ffedcdb064aad6ff33b6148725226cd608e
 F vsixtest/vsixtest_TemporaryKey.pfx e5b1b036facdb453873e7084e1cae9102ccc67a0
-P 0d2c992f3622a6272ca4a3caff6b21f619fe976b8df8b34eff066086f8df2202
-R 3144ef024c1b147facc59f065688290b
+P bfd5bf2c73110fcb36db9ba2a949ff516131fbd3e89325f88fe9f5c2b4ed87b2
+R aea18ad89757e208224db76c6eac01bc
+T *branch * as-materialize-redux
+T *sym-as-materialize-redux *
+T -sym-trunk *
 U drh
-Z 9a2772a6abeb20b71009a7863f85f821
+Z aa337b23bcebfb9b048b70a9fb74f61a

manifest.uuid

@@ -1 +1 @@
-bfd5bf2c73110fcb36db9ba2a949ff516131fbd3e89325f88fe9f5c2b4ed87b2
+ba59159fbe6b83fb6d79fbfee22d983768b0ebbaac7e99d2ac66c810e5e04100

src/delete.c

@@ -209,7 +209,11 @@ Expr *sqlite3LimitWhere(
   pSrc->a[0].pTab = 0;
   pSelectSrc = sqlite3SrcListDup(pParse->db, pSrc, 0);
   pSrc->a[0].pTab = pTab;
-  pSrc->a[0].pIBIndex = 0;
+  if( pSrc->a[0].fg.isIndexedBy ){
+    pSrc->a[0].u2.pIBIndex = 0;
+  }else if( pSrc->a[0].fg.isCte ){
+    pSrc->a[0].u2.pCteUse->nUse++;
+  }
 
   /* generate the SELECT expression tree. */
   pSelect = sqlite3SelectNew(pParse, pEList, pSelectSrc, pWhere, 0 ,0, 

src/expr.c

@@ -1544,7 +1544,10 @@ SrcList *sqlite3SrcListDup(sqlite3 *db, SrcList *p, int flags){
     if( pNewItem->fg.isIndexedBy ){
       pNewItem->u1.zIndexedBy = sqlite3DbStrDup(db, pOldItem->u1.zIndexedBy);
     }
-    pNewItem->pIBIndex = pOldItem->pIBIndex;
+    pNewItem->u2 = pOldItem->u2;
+    if( pNewItem->fg.isCte ){
+      pNewItem->u2.pCteUse->nUse++;
+    }
     if( pNewItem->fg.isTabFunc ){
       pNewItem->u1.pFuncArg = 
           sqlite3ExprListDup(db, pOldItem->u1.pFuncArg, flags);

src/prepare.c

@@ -597,7 +597,7 @@ void sqlite3ParserReset(Parse *pParse){
 **
 **       testcase( pParse->earlyCleanup );
 */
-void sqlite3ParserAddCleanup(
+void *sqlite3ParserAddCleanup(
   Parse *pParse,                      /* Destroy when this Parser finishes */
   void (*xCleanup)(sqlite3*,void*),   /* The cleanup routine */
   void *pPtr                          /* Pointer to object to be cleaned up */
@@ -610,10 +610,12 @@ void sqlite3ParserAddCleanup(
     pCleanup->xCleanup = xCleanup;
   }else{
     xCleanup(pParse->db, pPtr);
+    pPtr = 0;
 #if defined(SQLITE_DEBUG) || defined(SQLITE_COVERAGE_TEST)
     pParse->earlyCleanup = 1;
 #endif
   }
+  return pPtr;
 }
 
 /*

src/select.c

@@ -4726,7 +4726,7 @@ int sqlite3IndexedByLookup(Parse *pParse, SrcItem *pFrom){
     pParse->checkSchema = 1;
     return SQLITE_ERROR;
   }
-  pFrom->pIBIndex = pIdx;
+  pFrom->u2.pIBIndex = pIdx;
   return SQLITE_OK;
 }
 
@@ -4935,8 +4935,18 @@ static int resolveFromTermToCte(
     if( cannotBeFunction(pParse, pFrom) ) return 2;
 
     assert( pFrom->pTab==0 );
-    pFrom->pTab = pTab = sqlite3DbMallocZero(db, sizeof(Table));
+    pTab = sqlite3DbMallocZero(db, sizeof(Table));
     if( pTab==0 ) return 2;
+    if( pCte->pUse==0 ){
+      pCte->pUse = sqlite3DbMallocZero(db, sizeof(pCte->pUse[0]));
+      if( pCte->pUse==0
+       || sqlite3ParserAddCleanup(pParse,sqlite3DbFree,pCte->pUse)==0
+      ){
+        sqlite3DbFree(db, pTab);
+        return 2;
+      }
+    }
+    pFrom->pTab = pTab;
     pTab->nTabRef = 1;
     pTab->zName = sqlite3DbStrDup(db, pCte->zName);
     pTab->iPKey = -1;
@@ -4945,6 +4955,9 @@ static int resolveFromTermToCte(
     pFrom->pSelect = sqlite3SelectDup(db, pCte->pSelect, 0);
     if( db->mallocFailed ) return 2;
     assert( pFrom->pSelect );
+    pFrom->fg.isCte = 1;
+    pFrom->u2.pCteUse = pCte->pUse;
+    pCte->pUse->nUse++;
 
     /* Check if this is a recursive CTE. */
     pRecTerm = pSel = pFrom->pSelect;
@@ -6153,6 +6166,7 @@ int sqlite3Select(
   */
   for(i=0; i<pTabList->nSrc; i++){
     SrcItem *pItem = &pTabList->a[i];
+    SrcItem *pPrior;
     SelectDest dest;
     Select *pSub;
 #if !defined(SQLITE_OMIT_SUBQUERY) || !defined(SQLITE_OMIT_VIEW)
@@ -6212,6 +6226,7 @@ int sqlite3Select(
     ** inside the subquery.  This can help the subquery to run more efficiently.
     */
     if( OptimizationEnabled(db, SQLITE_PushDown)
+     && (pItem->fg.isCte==0 || pItem->u2.pCteUse->nUse<=1)
      && pushDownWhereTerms(pParse, pSub, p->pWhere, pItem->iCursor,
                            (pItem->fg.jointype & JT_OUTER)!=0)
     ){
@@ -6232,16 +6247,18 @@ int sqlite3Select(
 
     /* Generate code to implement the subquery
     **
-    ** The subquery is implemented as a co-routine if the subquery is
-    ** guaranteed to be the outer loop (so that it does not need to be
-    ** computed more than once)
+    ** The subquery is implemented as a co-routine if:
+    **    (1)  the subquery is guaranteed to be the outer loop (so that
+    **         it does not need to be computed more than once), and
+    **    (2)  the subquery is not a CTE that is used more then once.
     **
-    ** TODO: Are there other reasons beside (1) to use a co-routine
+    ** TODO: Are there other reasons beside (1) and (2) to use a co-routine
     ** implementation?
     */
     if( i==0
      && (pTabList->nSrc==1
             || (pTabList->a[1].fg.jointype&(JT_LEFT|JT_CROSS))!=0)  /* (1) */
+     && (pItem->fg.isCte==0 || pItem->u2.pCteUse->nUse<2)           /* (2) */
     ){
       /* Implement a co-routine that will return a single row of the result
       ** set on each invocation.
@@ -6261,16 +6278,30 @@ int sqlite3Select(
       sqlite3VdbeEndCoroutine(v, pItem->regReturn);
       sqlite3VdbeJumpHere(v, addrTop-1);
       sqlite3ClearTempRegCache(pParse);
+    }else if( pItem->fg.isCte && pItem->u2.pCteUse->addrM9e>0 ){
+      /* This is a CTE for which materialization code has already been
+      ** generated.  Invoke the subroutine to compute the materialization,
+      ** the make the pItem->iCursor be a copy of the ephemerial table that
+      ** holds the result of the materialization. */
+      CteUse *pCteUse = pItem->u2.pCteUse;
+      sqlite3VdbeAddOp2(v, OP_Gosub, pCteUse->regRtn, pCteUse->addrM9e);
+      if( pItem->iCursor!=pCteUse->iCur ){
+        sqlite3VdbeAddOp2(v, OP_OpenDup, pItem->iCursor, pCteUse->iCur);
+      }
+      pSub->nSelectRow = pCteUse->nRowEst;
+    }else if( (pPrior = isSelfJoinView(pTabList, pItem))!=0 ){
+      /* This view has already been materialized by a prior entry in
+      ** this same FROM clause.  Reuse it. */
+      if( pPrior->addrFillSub ){
+        sqlite3VdbeAddOp2(v, OP_Gosub, pPrior->regReturn, pPrior->addrFillSub);
+      }
+      sqlite3VdbeAddOp2(v, OP_OpenDup, pItem->iCursor, pPrior->iCursor);
+      pSub->nSelectRow = pPrior->pSelect->nSelectRow;
     }else{
-      /* Generate a subroutine that will fill an ephemeral table with
-      ** the content of this subquery.  pItem->addrFillSub will point
-      ** to the address of the generated subroutine.  pItem->regReturn
-      ** is a register allocated to hold the subroutine return address
-      */
+      /* Generate a subroutine that will materialize the view. */
       int topAddr;
       int onceAddr = 0;
       int retAddr;
-      SrcItem *pPrior;
 
       testcase( pItem->addrFillSub==0 ); /* Ticket c52b09c7f38903b1311 */
       pItem->regReturn = ++pParse->nMem;
@@ -6285,22 +6316,22 @@ int sqlite3Select(
       }else{
         VdbeNoopComment((v, "materialize \"%s\"", pItem->pTab->zName));
       }
-      pPrior = isSelfJoinView(pTabList, pItem);
-      if( pPrior ){
-        sqlite3VdbeAddOp2(v, OP_OpenDup, pItem->iCursor, pPrior->iCursor);
-        assert( pPrior->pSelect!=0 );
-        pSub->nSelectRow = pPrior->pSelect->nSelectRow;
-      }else{
       sqlite3SelectDestInit(&dest, SRT_EphemTab, pItem->iCursor);
       ExplainQueryPlan((pParse, 1, "MATERIALIZE %u", pSub->selId));
       sqlite3Select(pParse, pSub, &dest);
-      }
       pItem->pTab->nRowLogEst = pSub->nSelectRow;
       if( onceAddr ) sqlite3VdbeJumpHere(v, onceAddr);
       retAddr = sqlite3VdbeAddOp1(v, OP_Return, pItem->regReturn);
       VdbeComment((v, "end %s", pItem->pTab->zName));
       sqlite3VdbeChangeP1(v, topAddr, retAddr);
       sqlite3ClearTempRegCache(pParse);
+      if( pItem->fg.isCte ){
+        CteUse *pCteUse = pItem->u2.pCteUse;
+        pCteUse->addrM9e = pItem->addrFillSub;
+        pCteUse->regRtn = pItem->regReturn;
+        pCteUse->iCur = pItem->iCursor;
+        pCteUse->nRowEst = pSub->nSelectRow;
+      }
     }
     if( db->mallocFailed ) goto select_end;
     pParse->nHeight -= sqlite3SelectExprHeight(p);

src/sqliteInt.h

@@ -1137,6 +1137,7 @@ typedef struct Bitvec Bitvec;
 typedef struct CollSeq CollSeq;
 typedef struct Column Column;
 typedef struct Cte Cte;
+typedef struct CteUse CteUse;
 typedef struct Db Db;
 typedef struct DbFixer DbFixer;
 typedef struct Schema Schema;
@@ -2941,6 +2942,7 @@ struct SrcItem {
     unsigned viaCoroutine :1;  /* Implemented as a co-routine */
     unsigned isRecursive :1;   /* True for recursive reference in WITH */
     unsigned fromDDL :1;       /* Comes from sqlite_schema */
+    unsigned isCte :1;         /* This is a CTE */
   } fg;
   int iCursor;      /* The VDBE cursor number used to access this table */
   Expr *pOn;        /* The ON clause of a join */
@@ -2950,7 +2952,10 @@ struct SrcItem {
     char *zIndexedBy;    /* Identifier from "INDEXED BY <zIndex>" clause */
     ExprList *pFuncArg;  /* Arguments to table-valued-function */
   } u1;
+  union {
     Index *pIBIndex;  /* Index structure corresponding to u1.zIndexedBy */
+    CteUse *pCteUse;  /* CTE Usage info info fg.isCte is true */
+  } u2;
 };
 
 /*
@@ -3889,6 +3894,7 @@ struct Cte {
   ExprList *pCols;        /* List of explicit column names, or NULL */
   Select *pSelect;        /* The definition of this CTE */
   const char *zCteErr;    /* Error message for circular references */
+  CteUse *pUse;           /* Usage information for this CTE */
 };
 
 /*
@@ -3901,6 +3907,26 @@ struct With {
   Cte a[1];               /* For each CTE in the WITH clause.... */
 };
 
+/*
+** The Cte object is not guaranteed to persist for the entire duration
+** of code generation.  (The query flattener or other parser tree
+** edits might delete it.)  The following object records information
+** about each Common Table Expression that must be preserved for the
+** duration of the parse.
+**
+** The CteUse objects are freed using sqlite3ParserAddCleanup() rather
+** than sqlite3SelectDelete(), which is what enables them to persist
+** until the end of code generation.
+*/
+struct CteUse {
+  int nUse;              /* Number of users of this CTE */
+  int addrM9e;           /* Start of subroutine to compute materialization */
+  int regRtn;            /* Return address register for addrM9e subroutine */
+  int iCur;              /* Ephemeral table holding the materialization */
+  LogEst nRowEst;        /* Estimated number of rows in the table */
+};
+
+
 #ifdef SQLITE_DEBUG
 /*
 ** An instance of the TreeView object is used for printing the content of
@@ -4889,7 +4915,7 @@ sqlite3_int64 sqlite3StmtCurrentTime(sqlite3_context*);
 int sqlite3VdbeParameterIndex(Vdbe*, const char*, int);
 int sqlite3TransferBindings(sqlite3_stmt *, sqlite3_stmt *);
 void sqlite3ParserReset(Parse*);
-void sqlite3ParserAddCleanup(Parse*,void(*)(sqlite3*,void*),void*);
+void *sqlite3ParserAddCleanup(Parse*,void(*)(sqlite3*,void*),void*);
 #ifdef SQLITE_ENABLE_NORMALIZE
 char *sqlite3Normalize(Vdbe*, const char*);
 #endif

src/treeview.c

@@ -111,7 +111,10 @@ void sqlite3TreeViewWith(TreeView *pView, const With *pWith, u8 moreToFollow){
         }
         sqlite3_str_appendf(&x, ")");
       }
-      sqlite3_str_appendf(&x, " AS");
+      if( pCte->pUse ){
+        sqlite3_str_appendf(&x, " (pUse=0x%p, nUse=%d)", pCte->pUse,
+                 pCte->pUse->nUse);
+      }
       sqlite3StrAccumFinish(&x);
       sqlite3TreeViewItem(pView, zLine, i<pWith->nCte-1);
       sqlite3TreeViewSelect(pView, pCte->pSelect, 0);
@@ -150,6 +153,9 @@ void sqlite3TreeViewSrcList(TreeView *pView, const SrcList *pSrc){
     if( pItem->fg.fromDDL ){
       sqlite3_str_appendf(&x, " DDL");
     }
+    if( pItem->fg.isCte ){
+      sqlite3_str_appendf(&x, " CteUse=0x%p", pItem->u2.pCteUse);
+    }
     sqlite3StrAccumFinish(&x);
     sqlite3TreeViewItem(pView, zLine, i<pSrc->nSrc-1); 
     if( pItem->pSelect ){

src/where.c

@@ -2956,7 +2956,7 @@ static int whereLoopAddBtree(
 
   if( pSrc->fg.isIndexedBy ){
     /* An INDEXED BY clause specifies a particular index to use */
-    pProbe = pSrc->pIBIndex;
+    pProbe = pSrc->u2.pIBIndex;
   }else if( !HasRowid(pTab) ){
     pProbe = pTab->pIndex;
   }else{