Be more aggressive about reusing subqueries that appear on the RHS of IN

operators that have been replicated due to the predicate push-down optimization.

FossilOrigin-Name: 2accf32b6e45a396503c29eecc14a103bcc7b4c313cde921b26b489704060177

manifest

@@ -1,5 +1,5 @@
-C Add\scomment\susing\sthe\sname\s"predicate\spush-down\soptimization"\sto\swhat\swe\shave\nalso\scalled\s"WHERE-clause\spush\sdown".\s\sNo\schanges\sto\scode.
-D 2024-07-04T11:15:32.292
+C Be\smore\saggressive\sabout\sreusing\ssubqueries\sthat\sappear\son\sthe\sRHS\sof\sIN\noperators\sthat\shave\sbeen\sreplicated\sdue\sto\sthe\spredicate\spush-down\soptimization.
+D 2024-07-04T16:57:11.632
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
 F LICENSE.md df5091916dbb40e6e9686186587125e1b2ff51f022cc334e886c19a0e9982724
@@ -705,7 +705,7 @@ F src/date.c 13dd752847afb32ed70510ad7345a5b9c841f51ad904dba5d010f1fa3a6a324e
 F src/dbpage.c 80e46e1df623ec40486da7a5086cb723b0275a6e2a7b01d9f9b5da0f04ba2782
 F src/dbstat.c 3b677254d512fcafd4d0b341bf267b38b235ccfddbef24f9154e19360fa22e43
 F src/delete.c cb766727c78e715f9fb7ec8a7d03658ed2a3016343ca687acfcec9083cdca500
-F src/expr.c d6f614d0c90cf8d72bbd54dbae96040e7feb495a77991fdc1b1c67db68eeb326
+F src/expr.c 06c1d1a0f8b5025f001bc8219ac63fded7c33494a1863afdc78032ed1085ffef
 F src/fault.c 460f3e55994363812d9d60844b2a6de88826e007
 F src/fkey.c 852f93c0ef995e0c2b8983059a2b97151c194cc8259e21f5bc2b7ac508348c2a
 F src/func.c 1f61e32e7a357e615b5d2e774bee563761fce4f2fd97ecb0f72c33e62a2ada5f
@@ -826,10 +826,10 @@ F src/utf.c f23165685a67b4caf8ec08fb274cb3f319103decfb2a980b7cfd55d18dfa855e
 F src/util.c 5d1a0134cf4240648d1c6bb5cc8efaca0ea2b5d5c840985aec7e947271f04375
 F src/vacuum.c 604fcdaebe76f3497c855afcbf91b8fa5046b32de3045bab89cc008d68e40104
 F src/vdbe.c f941fd8dbd1a9e073866a941e7fbb553f1de2d25b249611005d8c27d5eac8995
-F src/vdbe.h c2d78d15112c3fc5ab87f5e8e0b75d2db1c624409de2e858c3d1aafb1650bb4f
+F src/vdbe.h c2549a215898a390de6669cfa32adba56f0d7e17ba5a7f7b14506d6fd5f0c36a
 F src/vdbeInt.h 949669dfd8a41550d27dcb905b494f2ccde9a2e6c1b0b04daa1227e2e74c2b2c
 F src/vdbeapi.c 80235ac380e9467fec1cb0883354d841f2a771976e766995f7e0c77f845406df
-F src/vdbeaux.c 6e37cb918506c28fe7657454fcbc2e01e66bfba4164f306c2f075fd5c5fef609
+F src/vdbeaux.c 8d522d012c0545035be52e46b4cc5dfa752e376e17137816e558b3ab6bec9e5d
 F src/vdbeblob.c 13f9287b55b6356b4b1845410382d6bede203ceb29ef69388a4a3d007ffacbe5
 F src/vdbemem.c 831a244831eaa45335f9ae276b50a7a82ee10d8c46c2c72492d4eb8c98d94d89
 F src/vdbesort.c 237840ca1947511fa59bd4e18b9eeae93f2af2468c34d2427b059f896230a547
@@ -1518,7 +1518,7 @@ F test/printf.test 685fec5a0c5af2490ab0632775a301554361d674211d690f5bee0a97b0533
 F test/printf2.test 3f55c1871a5a65507416076f6eb97e738d5210aeda7595a74ee895f2224cce60
 F test/progress.test ebab27f670bd0d4eb9d20d49cef96e68141d92fb
 F test/ptrchng.test ef1aa72d6cf35a2bbd0869a649b744e9d84977fc
-F test/pushdown.test 9e655df51bc6559608dcc7af89a36f727eff520b4860261ed15a5f95a5ebbcd8
+F test/pushdown.test 84b525767442b3695d671f9df59dd91cf0ed8fb24cbbcdc55959f0dadeee8b39
 F test/queryonly.test 5f653159e0f552f0552d43259890c1089391dcca
 F test/quick.test 1681febc928d686362d50057c642f77a02c62e57
 F test/quickcheck.test a4b7e878cd97e46108291c409b0bf8214f29e18fddd68a42bc5c1375ad1fb80a
@@ -2195,8 +2195,11 @@ F vsixtest/vsixtest.tcl 6a9a6ab600c25a91a7acc6293828957a386a8a93
 F vsixtest/vsixtest.vcxproj.data 2ed517e100c66dc455b492e1a33350c1b20fbcdc
 F vsixtest/vsixtest.vcxproj.filters 37e51ffedcdb064aad6ff33b6148725226cd608e
 F vsixtest/vsixtest_TemporaryKey.pfx e5b1b036facdb453873e7084e1cae9102ccc67a0
-P a204ffc06b468c2edf8f11ccf3de639edd4f8282e69a44ceeb68a4d3a43e77ea
-R 7ddaa7e2074acfdf6d0130f12e8505c2
+P be77fe701c7b7d3407800e1ed6a4c8be14035831790419cc99362bff41c65b60
+R 53a737f2af23a108773b1a9f30c846b8
+T *branch * reuse-subqueries
+T *sym-reuse-subqueries *
+T -sym-trunk *
 U drh
-Z 940dec88e4e92fbd5a212ce2f1e6eedc
+Z 375183e5bacb7ab157909f081f98b1e8
 # Remove this line to create a well-formed Fossil manifest.

manifest.uuid

@@ -1 +1 @@
-be77fe701c7b7d3407800e1ed6a4c8be14035831790419cc99362bff41c65b60
+2accf32b6e45a396503c29eecc14a103bcc7b4c313cde921b26b489704060177

src/expr.c

@@ -3420,6 +3420,46 @@ void sqlite3VectorErrorMsg(Parse *pParse, Expr *pExpr){
   }
 }
 
+#ifndef SQLITE_OMIT_SUBQUERY
+/*
+** Scan all previously generated bytecode looking for an OP_BeginSubrtn
+** that is compatible with pExpr.  If found, add the y.sub values
+** to pExpr and return true.  If not found, return false.
+*/
+static int findCompatibleInRhsSubrtn(
+  Parse *pParse,          /* Parsing context */
+  Expr *pExpr,            /* IN operator with RHS that we want to reuse */
+  SubrtnSig *pNewSig      /* Signature for the IN operator */
+){
+  VdbeOp *pOp, *pEnd;
+  SubrtnSig *pSig;
+  Vdbe *v;
+
+  if( pNewSig==0 ) return 0;
+  assert( pExpr->op==TK_IN );
+  assert( !ExprUseYSub(pExpr) );
+  assert( ExprUseXSelect(pExpr) );
+  v = pParse->pVdbe;
+  assert( v!=0 );
+  pOp = sqlite3VdbeGetOp(v, 1);
+  pEnd = sqlite3VdbeGetLastOp(v);
+  for(; pOp<pEnd; pOp++){
+    if( pOp->opcode!=OP_BeginSubrtn ) continue;
+    if( pOp->p4type!=P4_SUBRTNSIG ) continue;
+    pSig = pOp->p4.pSubrtnSig;
+    assert( pSig!=0 );
+    if( pNewSig->selId!=pSig->selId ) continue;
+    if( strcmp(pNewSig->zAff,pSig->zAff)!=0 ) continue;
+    pExpr->y.sub.iAddr = pSig->iAddr;
+    pExpr->y.sub.regReturn = pSig->regReturn;
+    pExpr->iTable = pSig->iTable;
+    ExprSetProperty(pExpr, EP_Subrtn);
+    return 1;
+  }
+  return 0;
+}
+#endif /* SQLITE_OMIT_SUBQUERY */
+
 #ifndef SQLITE_OMIT_SUBQUERY
 /*
 ** Generate code that will construct an ephemeral table containing all terms
@@ -3469,11 +3509,30 @@ void sqlite3CodeRhsOfIN(
   ** and reuse it many names.
   */
   if( !ExprHasProperty(pExpr, EP_VarSelect) && pParse->iSelfTab==0 ){
-    /* Reuse of the RHS is allowed */
-    /* If this routine has already been coded, but the previous code
-    ** might not have been invoked yet, so invoke it now as a subroutine.
+    /* Reuse of the RHS is allowed
+    **
+    ** Compute a signature for the RHS of the IN operator to facility
+    ** finding and reusing prior instances of the same IN operator.
     */
-    if( ExprHasProperty(pExpr, EP_Subrtn) ){
+    SubrtnSig *pSig;
+    if( !ExprUseXSelect(pExpr) ){
+      pSig = 0;
+    }else{
+      assert( pExpr->x.pSelect!=0 );
+      pSig = sqlite3DbMallocRawNN(pParse->db, sizeof(pSig[0]));
+      if( pSig ){
+        pSig->selId = pExpr->x.pSelect->selId;
+        pSig->zAff = exprINAffinity(pParse, pExpr);
+      }
+    }
+
+    /* Check to see if there is a prior materialization of the RHS of
+    ** this IN operator.  If there is, then make use of that prior
+    ** materialization rather than recomputing it.
+    */
+    if( ExprHasProperty(pExpr, EP_Subrtn) 
+     || findCompatibleInRhsSubrtn(pParse, pExpr, pSig)
+    ){
       addrOnce = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v);
       if( ExprUseXSelect(pExpr) ){
         ExplainQueryPlan((pParse, 0, "REUSE LIST SUBQUERY %d",
@@ -3485,6 +3544,10 @@ void sqlite3CodeRhsOfIN(
       assert( iTab!=pExpr->iTable );
       sqlite3VdbeAddOp2(v, OP_OpenDup, iTab, pExpr->iTable);
       sqlite3VdbeJumpHere(v, addrOnce);
+      if( pSig ){
+        sqlite3DbFree(pParse->db, pSig->zAff);
+        sqlite3DbFree(pParse->db, pSig);
+      }
       return;
     }
 
@@ -3495,7 +3558,12 @@ void sqlite3CodeRhsOfIN(
     pExpr->y.sub.regReturn = ++pParse->nMem;
     pExpr->y.sub.iAddr =
       sqlite3VdbeAddOp2(v, OP_BeginSubrtn, 0, pExpr->y.sub.regReturn) + 1;
-
+    if( pSig ){
+      pSig->iAddr = pExpr->y.sub.iAddr;
+      pSig->regReturn = pExpr->y.sub.regReturn;
+      pSig->iTable = iTab;
+      sqlite3VdbeChangeP4(v, -1, (const char*)pSig, P4_SUBRTNSIG);
+    }
     addrOnce = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v);
   }
 

src/vdbe.h

@@ -32,6 +32,19 @@ typedef struct Vdbe Vdbe;
 */
 typedef struct sqlite3_value Mem;
 typedef struct SubProgram SubProgram;
+typedef struct SubrtnSig SubrtnSig;
+
+/*
+** A signature for a reusable subroutine that materializes the RHS of
+** an IN operator.
+*/
+struct SubrtnSig {
+  int selId;          /* SELECT-id for the SELECT statement on the RHS */
+  char *zAff;         /* Affinity of the overall IN expression */
+  int iTable;         /* Ephemeral table generated by the subroutine */
+  int iAddr;          /* Subroutine entry address */
+  int regReturn;      /* Register used to hold return address */
+};
 
 /*
 ** A single instruction of the virtual machine has an opcode
@@ -60,6 +73,7 @@ struct VdbeOp {
     u32 *ai;               /* Used when p4type is P4_INTARRAY */
     SubProgram *pProgram;  /* Used when p4type is P4_SUBPROGRAM */
     Table *pTab;           /* Used when p4type is P4_TABLE */
+    SubrtnSig *pSubrtnSig; /* Used when p4type is P4_SUBRTNSIG */
 #ifdef SQLITE_ENABLE_CURSOR_HINTS
     Expr *pExpr;           /* Used when p4type is P4_EXPR */
 #endif
@@ -127,6 +141,7 @@ typedef struct VdbeOpList VdbeOpList;
 #define P4_INTARRAY   (-14) /* P4 is a vector of 32-bit integers */
 #define P4_FUNCCTX    (-15) /* P4 is a pointer to an sqlite3_context object */
 #define P4_TABLEREF   (-16) /* Like P4_TABLE, but reference counted */
+#define P4_SUBRTNSIG  (-17) /* P4 is a SubrtnSig pointer */
 
 /* Error message codes for OP_Halt */
 #define P5_ConstraintNotNull 1

src/vdbeaux.c

@@ -1413,6 +1413,12 @@ static void freeP4(sqlite3 *db, int p4type, void *p4){
       if( db->pnBytesFreed==0 ) sqlite3DeleteTable(db, (Table*)p4);
       break;
     }
+    case P4_SUBRTNSIG: {
+      SubrtnSig *pSig = (SubrtnSig*)p4;
+      sqlite3DbFree(db, pSig->zAff);
+      sqlite3DbFree(db, pSig);
+      break;
+    }
   }
 }
 
@@ -1992,6 +1998,9 @@ char *sqlite3VdbeDisplayP4(sqlite3 *db, Op *pOp){
       zP4 = pOp->p4.pTab->zName;
       break;
     }
+    case P4_SUBRTNSIG: {
+      break;
+    }
     default: {
       zP4 = pOp->p4.z;
     }

test/pushdown.test

@@ -279,13 +279,9 @@ do_eqp_test 6.1 {
   |     |     `--CREATE BLOOM FILTER
   |     `--UNION ALL
   |        |--SEARCH t02 USING INDEX t02x (w=? AND x=? AND y>? AND y<?)
-  |        `--LIST SUBQUERY xxxxxx
-  |           |--SCAN k
-  |           `--CREATE BLOOM FILTER
+  |        `--REUSE LIST SUBQUERY xxxxxx
   |--SEARCH t0
-  `--LIST SUBQUERY xxxxxx
-     |--SCAN k
-     `--CREATE BLOOM FILTER
+  `--REUSE LIST SUBQUERY xxxxxx
 }
 # ^^^^--- The key feature above is that the SEARCH for each subquery
 # uses all three fields of the index w, x, and y.  Prior to the push-down
@@ -307,17 +303,9 @@ do_eqp_test 6.2 {
   |     |     `--CREATE BLOOM FILTER
   |     `--UNION ALL
   |        |--SEARCH t02 USING INDEX t02x (w=? AND x=? AND y>? AND y<?)
-  |        `--LIST SUBQUERY xxxxxx
-  |           |--CO-ROUTINE v1
-  |           |  `--SCAN 3 CONSTANT ROWS
-  |           |--SCAN v1
-  |           `--CREATE BLOOM FILTER
+  |        `--REUSE LIST SUBQUERY xxxxxx
   |--SEARCH t0
-  `--LIST SUBQUERY xxxxxx
-     |--CO-ROUTINE v1
-     |  `--SCAN 3 CONSTANT ROWS
-     |--SCAN v1
-     `--CREATE BLOOM FILTER
+  `--REUSE LIST SUBQUERY xxxxxx
 }
 do_eqp_test 6.3 {
   SELECT max(z) FROM t0 WHERE w=123 AND x IN k1 AND y BETWEEN 44 AND 55;
@@ -332,13 +320,9 @@ do_eqp_test 6.3 {
   |     |     `--CREATE BLOOM FILTER
   |     `--UNION ALL
   |        |--SEARCH t02 USING INDEX t02x (w=? AND x=? AND y>? AND y<?)
-  |        `--LIST SUBQUERY xxxxxx
-  |           |--SCAN k1
-  |           `--CREATE BLOOM FILTER
+  |        `--REUSE LIST SUBQUERY xxxxxx
   |--SEARCH t0
-  `--LIST SUBQUERY xxxxxx
-     |--SCAN k1
-     `--CREATE BLOOM FILTER
+  `--REUSE LIST SUBQUERY xxxxxx
 }
 
 finish_test