diff --git a/manifest b/manifest index 4ed8e1a4b9..9513f2ff2e 100644 --- a/manifest +++ b/manifest @@ -1,5 +1,5 @@ -C Use\sOP_SorterOpen\sinstead\sof\sOP_OpenEphemeral\sto\simplement\sGROUP\sBY. -D 2011-09-01T16:01:27.777 +C Instead\sof\sa\stemporary\sb-tree,\suse\sa\slinked-list\sand\smerge-sort\sto\ssort\srecords\sin\smain\smemory\sin\svdbesort.c. +D 2011-09-02T10:31:11.173 F Makefile.arm-wince-mingw32ce-gcc d6df77f1f48d690bd73162294bbba7f59507c72f F Makefile.in d314143fa6be24828021d3f583ad37d9afdce505 F Makefile.linux-gcc 91d710bdc4998cb015f39edf3cb314ec4f4d7e23 @@ -124,16 +124,16 @@ F src/auth.c 523da7fb4979469955d822ff9298352d6b31de34 F src/backup.c 28a4fe55327ff708bfaf9d4326d02686f7a553c3 F src/bitvec.c af50f1c8c0ff54d6bdb7a80e2fceca5a93670bef F src/btmutex.c 976f45a12e37293e32cae0281b15a21d48a8aaa7 -F src/btree.c 4a2856b3bde9959986a7b9327841b3ff94023784 -F src/btree.h 9ddf04226eac592d4cc3709c5a8b33b2351ff5f7 +F src/btree.c bc2099e7d3c22c52b2c54349b9c07c04f2a810d0 +F src/btree.h f5d775cd6cfc7ac32a2535b70e8d2af48ef5f2ce F src/btreeInt.h 67978c014fa4f7cc874032dd3aacadd8db656bc3 -F src/build.c 2d5de52df616a3bf5a659cbca85211c46e2ba9bd +F src/build.c dc367138cb3625e6d42b389e05d7267aece5753c F src/callback.c 0425c6320730e6d3981acfb9202c1bed9016ad1a F src/complete.c dc1d136c0feee03c2f7550bafc0d29075e36deac F src/ctime.c e3132ec65240b2e2f3d50831021eac387f27584d F src/date.c a3c6842bad7ae632281811de112a8ba63ff08ab3 F src/delete.c ff68e5ef23aee08c0ff528f699a19397ed8bbed8 -F src/expr.c 4bbdfaf66bc614be9254ce0c26a17429067a3e07 +F src/expr.c cbcd8c2f1588a9862291a081699854c5e1cb28ab F src/fault.c 160a0c015b6c2629d3899ed2daf63d75754a32bb F src/fkey.c 9f00ea98f6b360d477b5a78b5b59a1fbde82431c F src/func.c 59bb046d7e3df1ab512ac339ccb0a6f996a17cb7 @@ -179,11 +179,11 @@ F src/printf.c 585a36b6a963df832cfb69505afa3a34ed5ef8a1 F src/random.c cd4a67b3953b88019f8cd4ccd81394a8ddfaba50 F src/resolve.c 36368f44569208fa074e61f4dd0b6c4fb60ca2b4 F src/rowset.c 69afa95a97c524ba6faf3805e717b5b7ae85a697 -F src/select.c 037ee5501fe0e743fa98936902c200ed9ed69156 +F src/select.c 32d0f4e5513362706b8973e7f1b87cd0885dfbf5 F src/shell.c bbe7818ff5bc8614105ceb81ad67b8bdc0b671dd F src/sqlite.h.in 0a6c9c23337fd1352c5c75a613ff9533aa7d91cb F src/sqlite3ext.h 1a1a4f784aa9c3b00edd287940197de52487cd93 -F src/sqliteInt.h f6debf9a9eb8463ab2ef8be4b2b740ea9af5afba +F src/sqliteInt.h 723cda73a33c91f5a0a145f4c0ced45d94921079 F src/sqliteLimit.h 164b0e6749d31e0daa1a4589a169d31c0dec7b3d F src/status.c 7ac64842c86cec2fc1a1d0e5c16d3beb8ad332bf F src/table.c 2cd62736f845d82200acfa1287e33feb3c15d62e @@ -238,14 +238,14 @@ F src/update.c 74a6cfb34e9732c1e2a86278b229913b4b51eeec F src/utf.c c53eb7404b3eb5c1cbb5655c6a7a0e0ce6bd50f0 F src/util.c 06302ffd2b80408d4f6c7af71f7090e0cf8d8ff7 F src/vacuum.c 05513dca036a1e7848fe18d5ed1265ac0b32365e -F src/vdbe.c 9260e5138855399bea2611a29da336688bfa1b79 +F src/vdbe.c da9c7efc48dc79d7785f3f17a1c3df514bf18489 F src/vdbe.h c1eeedacab6bcf1e7c2cf8203ba9763a616f9a86 -F src/vdbeInt.h 51a902e12c7d571e3b516e5407e30f996494aafe +F src/vdbeInt.h a255da14be8c2794ce38e0d2142877bb29df9105 F src/vdbeapi.c 11dc47987abacb76ad016dcf5abc0dc422482a98 F src/vdbeaux.c e58acbc5ea3823922a0cd8fa21f94f39af51ee88 F src/vdbeblob.c f024f0bf420f36b070143c32b15cc7287341ffd3 F src/vdbemem.c 5e6effb96dd53d233361cbfaa3f0a43b9af689e9 -F src/vdbesort.c f3d043a1bab7409d4a23cd7a35287c3ac440a167 +F src/vdbesort.c 9c2e8ca23c2413a5162a4433cb72377588d896f8 F src/vdbetrace.c 5d0dc3d5fd54878cc8d6d28eb41deb8d5885b114 F src/vtab.c 901791a47318c0562cd0c676a2c6ff1bc530e582 F src/wal.c 3154756177d6219e233d84291d5b05f4e06ff5e9 @@ -511,7 +511,7 @@ F test/incrvacuum_ioerr.test 22f208d01c528403240e05beecc41dc98ed01637 F test/index.test b5429732b3b983fa810e3ac867d7ca85dae35097 F test/index2.test ee83c6b5e3173a3d7137140d945d9a5d4fdfb9d6 F test/index3.test 423a25c789fc8cc51aaf2a4370bbdde2d9e9eed7 -F test/index4.test c82a59c9ae2ac01804bdb100162dca057318f40f +F test/index4.test 2983216eb8c86ee62d9ed7cb206b5cc3331c0026 F test/indexedby.test be501e381b82b2f8ab406309ba7aac46e221f4ad F test/indexfault.test 31d4ab9a7d2f6e9616933eb079722362a883eb1d F test/init.test 15c823093fdabbf7b531fe22cf037134d09587a7 @@ -961,7 +961,7 @@ F tool/symbols.sh caaf6ccc7300fd43353318b44524853e222557d5 F tool/tostr.awk e75472c2f98dd76e06b8c9c1367f4ab07e122d06 F tool/vdbe-compress.tcl d70ea6d8a19e3571d7ab8c9b75cba86d1173ff0f F tool/warnings.sh b7fdb2cc525f5ef4fa43c80e771636dd3690f9d2 -P bab2e560f6cb989c83a96aad60f666960ede7abe -R 47ae22012e17b717f429ee6a9d305e96 -U drh -Z 4edda214a8d1e652ef01c1f98ad8fd39 +P ebf819aaa555bd79fddfc0a6f9827a2539095d6c +R c9d892df81c49366e067933871c3b776 +U dan +Z b3581bd03fb5a181801f57c568eb418b diff --git a/manifest.uuid b/manifest.uuid index e5664d914b..f82002ddab 100644 --- a/manifest.uuid +++ b/manifest.uuid @@ -1 +1 @@ -ebf819aaa555bd79fddfc0a6f9827a2539095d6c \ No newline at end of file +7769fb988d9be0f2d8129aaac19620ac88f9b4a6 \ No newline at end of file diff --git a/src/btree.c b/src/btree.c index 7166b93b90..b335579bd0 100644 --- a/src/btree.c +++ b/src/btree.c @@ -1734,22 +1734,11 @@ int sqlite3BtreeOpen( /* A BTREE_SINGLE database is always a temporary and/or ephemeral */ assert( (flags & BTREE_SINGLE)==0 || isTempDb ); - /* The BTREE_SORTER flag is only used if SQLITE_OMIT_MERGE_SORT is undef */ -#ifdef SQLITE_OMIT_MERGE_SORT - assert( (flags & BTREE_SORTER)==0 ); -#endif - - /* BTREE_SORTER is always on a BTREE_SINGLE, BTREE_OMIT_JOURNAL */ - assert( (flags & BTREE_SORTER)==0 || - (flags & (BTREE_SINGLE|BTREE_OMIT_JOURNAL)) - ==(BTREE_SINGLE|BTREE_OMIT_JOURNAL) ); - if( db->flags & SQLITE_NoReadlock ){ flags |= BTREE_NO_READLOCK; } if( isMemdb ){ flags |= BTREE_MEMORY; - flags &= ~BTREE_SORTER; } if( (vfsFlags & SQLITE_OPEN_MAIN_DB)!=0 && (isMemdb || isTempDb) ){ vfsFlags = (vfsFlags & ~SQLITE_OPEN_MAIN_DB) | SQLITE_OPEN_TEMP_DB; diff --git a/src/btree.h b/src/btree.h index ce19826ad8..9e3a73b3b6 100644 --- a/src/btree.h +++ b/src/btree.h @@ -61,7 +61,6 @@ int sqlite3BtreeOpen( #define BTREE_MEMORY 4 /* This is an in-memory DB */ #define BTREE_SINGLE 8 /* The file contains at most 1 b-tree */ #define BTREE_UNORDERED 16 /* Use of a hash implementation is OK */ -#define BTREE_SORTER 32 /* Used as accumulator in external merge sort */ int sqlite3BtreeClose(Btree*); int sqlite3BtreeSetCacheSize(Btree*,int); diff --git a/src/build.c b/src/build.c index 29fbf92713..0de1020b9b 100644 --- a/src/build.c +++ b/src/build.c @@ -2326,6 +2326,7 @@ static void sqlite3RefillIndex(Parse *pParse, Index *pIndex, int memRootPage){ int iIdx = pParse->nTab++; /* Btree cursor used for pIndex */ int iSorter = iTab; /* Cursor opened by OpenSorter (if in use) */ int addr1; /* Address of top of loop */ + int addr2; /* Address to jump to for next iteration */ int tnum; /* Root page of index */ Vdbe *v; /* Generate code into this virtual machine */ KeyInfo *pKey; /* KeyInfo for index */ @@ -2372,25 +2373,34 @@ static void sqlite3RefillIndex(Parse *pParse, Index *pIndex, int memRootPage){ if( bUseSorter ){ iSorter = pParse->nTab++; sqlite3VdbeAddOp4(v, OP_OpenSorter, iSorter, 0, 0, (char*)pKey, P4_KEYINFO); - sqlite3VdbeChangeP5(v, BTREE_SORTER); } /* Open the table. Loop through all rows of the table, inserting index ** records into the sorter. */ sqlite3OpenTable(pParse, iTab, iDb, pTab, OP_OpenRead); addr1 = sqlite3VdbeAddOp2(v, OP_Rewind, iTab, 0); + addr2 = addr1 + 1; regRecord = sqlite3GetTempReg(pParse); regIdxKey = sqlite3GenerateIndexKey(pParse, pIndex, iTab, regRecord, 1); if( bUseSorter ){ - sqlite3VdbeAddOp2(v, OP_IdxInsert, iSorter, regRecord); + sqlite3VdbeAddOp2(v, OP_SorterInsert, iSorter, regRecord); sqlite3VdbeAddOp2(v, OP_Next, iTab, addr1+1); sqlite3VdbeJumpHere(v, addr1); - addr1 = sqlite3VdbeAddOp2(v, OP_Sort, iSorter, 0); - sqlite3VdbeAddOp2(v, OP_RowKey, iSorter, regRecord); - } - - if( pIndex->onError!=OE_None ){ + addr1 = sqlite3VdbeAddOp2(v, OP_SorterSort, iSorter, 0); + if( pIndex->onError!=OE_None ){ + int j2 = sqlite3VdbeCurrentAddr(v) + 3; + sqlite3VdbeAddOp2(v, OP_Goto, 0, j2); + addr2 = sqlite3VdbeCurrentAddr(v); + sqlite3VdbeAddOp3(v, OP_SorterCompare, iSorter, j2, regRecord); + sqlite3HaltConstraint( + pParse, OE_Abort, "indexed columns are not unique", P4_STATIC + ); + }else{ + addr2 = sqlite3VdbeCurrentAddr(v); + } + sqlite3VdbeAddOp2(v, OP_SorterData, iSorter, regRecord); + }else if( pIndex->onError!=OE_None ){ const int regRowid = regIdxKey + pIndex->nColumn; const int j2 = sqlite3VdbeCurrentAddr(v) + 2; void * const pRegKey = SQLITE_INT_TO_PTR(regIdxKey); @@ -2411,7 +2421,7 @@ static void sqlite3RefillIndex(Parse *pParse, Index *pIndex, int memRootPage){ sqlite3VdbeAddOp3(v, OP_IdxInsert, iIdx, regRecord, bUseSorter); sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); sqlite3ReleaseTempReg(pParse, regRecord); - sqlite3VdbeAddOp2(v, OP_Next, iSorter, addr1+1); + sqlite3VdbeAddOp2(v, bUseSorter ? OP_SorterNext : OP_Next, iSorter, addr2); sqlite3VdbeJumpHere(v, addr1); sqlite3VdbeAddOp1(v, OP_Close, iTab); diff --git a/src/expr.c b/src/expr.c index ab4547db9c..ab218078db 100644 --- a/src/expr.c +++ b/src/expr.c @@ -2287,7 +2287,7 @@ int sqlite3ExprCodeTarget(Parse *pParse, Expr *pExpr, int target){ inReg = pCol->iMem; break; }else if( pAggInfo->useSortingIdx ){ - sqlite3VdbeAddOp3(v, OP_Column, pAggInfo->sortingIdx, + sqlite3VdbeAddOp3(v, OP_Column, pAggInfo->sortingIdxPTab, pCol->iSorterColumn, target); break; } diff --git a/src/select.c b/src/select.c index d8332eaaea..7d9d26602e 100644 --- a/src/select.c +++ b/src/select.c @@ -4186,7 +4186,7 @@ int sqlite3Select( sqlite3ReleaseTempReg(pParse, regRecord); sqlite3ReleaseTempRange(pParse, regBase, nCol); sqlite3WhereEnd(pWInfo); - sortPTab = pParse->nTab++; + sAggInfo.sortingIdxPTab = sortPTab = pParse->nTab++; sortOut = sqlite3GetTempReg(pParse); sqlite3VdbeAddOp3(v, OP_OpenPseudo, sortPTab, sortOut, nCol); sqlite3VdbeAddOp2(v, OP_SorterSort, sAggInfo.sortingIdx, addrEnd); diff --git a/src/sqliteInt.h b/src/sqliteInt.h index 18beb32e0f..694f4694eb 100644 --- a/src/sqliteInt.h +++ b/src/sqliteInt.h @@ -1550,6 +1550,7 @@ struct AggInfo { u8 useSortingIdx; /* In direct mode, reference the sorting index rather ** than the source table */ int sortingIdx; /* Cursor number of the sorting index */ + int sortingIdxPTab; /* Cursor number of pseudo-table */ ExprList *pGroupBy; /* The group by clause */ int nSortingColumn; /* Number of columns in the sorting index */ struct AggInfo_col { /* For each column used in source tables */ diff --git a/src/vdbe.c b/src/vdbe.c index f4ae4c44a6..d9310e0563 100644 --- a/src/vdbe.c +++ b/src/vdbe.c @@ -3162,15 +3162,7 @@ case OP_OpenWrite: { ** by this opcode will be used for automatically created transient ** indices in joins. */ -/* Opcode: OpenSorter P1 P2 * P4 * -** -** This opcode works like OP_OpenEphemeral except that it opens -** a transient index that is specifically designed to sort large -** tables using an external merge-sort algorithm. -*/ -case OP_OpenSorter: case OP_OpenAutoindex: -case OP_SorterOpen: case OP_OpenEphemeral: { VdbeCursor *pCx; static const int vfsFlags = @@ -3181,7 +3173,6 @@ case OP_OpenEphemeral: { SQLITE_OPEN_TRANSIENT_DB; assert( pOp->p1>=0 ); - assert( (pOp->opcode==OP_OpenSorter)==((pOp->p5 & BTREE_SORTER)!=0) ); pCx = allocateCursor(p, pOp->p1, pOp->p2, -1, 1); if( pCx==0 ) goto no_mem; pCx->nullRow = 1; @@ -3215,12 +3206,24 @@ case OP_OpenEphemeral: { } pCx->isOrdered = (pOp->p5!=BTREE_UNORDERED); pCx->isIndex = !pCx->isTable; - pCx->isSorter = pOp->opcode==OP_SorterOpen; -#ifndef SQLITE_OMIT_MERGE_SORT - if( rc==SQLITE_OK && pOp->opcode==OP_OpenSorter ){ - rc = sqlite3VdbeSorterInit(db, pCx); - } -#endif + break; +} + +/* Opcode: OpenSorter P1 P2 * P4 * +** +** This opcode works like OP_OpenEphemeral except that it opens +** a transient index that is specifically designed to sort large +** tables using an external merge-sort algorithm. +*/ +case OP_SorterOpen: +case OP_OpenSorter: { + VdbeCursor *pCx; + pCx = allocateCursor(p, pOp->p1, pOp->p2, -1, 1); + if( pCx==0 ) goto no_mem; + pCx->pKeyInfo = pOp->p4.pKeyInfo; + pCx->pKeyInfo->enc = ENC(p->db); + pCx->isSorter = 1; + rc = sqlite3VdbeSorterInit(db, pCx); break; } @@ -4072,6 +4075,40 @@ case OP_ResetCount: { break; } +/* Opcode: SorterCompare P1 P2 P3 +** +** P1 is a sorter cursor. This instruction compares the record blob in +** register P3 with the entry that the sorter cursor currently points to. +** If, excluding the rowid fields at the end, the two records are a match, +** fall through to the next instruction. Otherwise, jump to instruction P2. +*/ +case OP_SorterCompare: { + VdbeCursor *pC; + int res; + + pC = p->apCsr[pOp->p1]; + assert( isSorter(pC) ); + pIn3 = &aMem[pOp->p3]; + rc = sqlite3VdbeSorterCompare(pC, pIn3, &res); + if( res ){ + pc = pOp->p2-1; + } + break; +}; + +/* Opcode: SorterData P1 P2 * * * +** +** Write into register P2 the current sorter data for sorter cursor P1. +*/ +case OP_SorterData: { + VdbeCursor *pC; + pOut = &aMem[pOp->p2]; + pC = p->apCsr[pOp->p1]; + assert( pC->isSorter ); + rc = sqlite3VdbeSorterRowkey(pC, pOut); + break; +} + /* Opcode: RowData P1 P2 * * * ** ** Write into register P2 the complete row data for cursor P1. @@ -4092,7 +4129,6 @@ case OP_ResetCount: { ** If the P1 cursor must be pointing to a valid row (not a NULL row) ** of a real table, not a pseudo-table. */ -case OP_SorterData: case OP_RowKey: case OP_RowData: { VdbeCursor *pC; @@ -4106,6 +4142,7 @@ case OP_RowData: { /* Note that RowKey and RowData are really exactly the same instruction */ assert( pOp->p1>=0 && pOp->p1nCursor ); pC = p->apCsr[pOp->p1]; + assert( pC->isSorter==0 ); assert( pC->isTable || pOp->opcode!=OP_RowData ); assert( pC->isIndex || pOp->opcode==OP_RowData ); assert( pC!=0 ); @@ -4113,12 +4150,6 @@ case OP_RowData: { assert( pC->pseudoTableReg==0 ); assert( pC->isSorter==(pOp->opcode==OP_SorterData) ); - if( isSorter(pC) ){ - assert( pOp->opcode==OP_RowKey ); - rc = sqlite3VdbeSorterRowkey(pC, pOut); - break; - } - assert( pC->pCursor!=0 ); pCrsr = pC->pCursor; assert( sqlite3BtreeCursorIsValid(pCrsr) ); @@ -4368,7 +4399,7 @@ case OP_Next: { /* jump */ } assert( pC->isSorter==(pOp->opcode==OP_SorterNext) ); if( isSorter(pC) ){ - assert( pOp->opcode==OP_Next ); + assert( pOp->opcode==OP_SorterNext ); rc = sqlite3VdbeSorterNext(db, pC, &res); }else{ res = 1; @@ -4421,16 +4452,17 @@ case OP_IdxInsert: { /* in2 */ assert( pC->isTable==0 ); rc = ExpandBlob(pIn2); if( rc==SQLITE_OK ){ - nKey = pIn2->n; - zKey = pIn2->z; - rc = sqlite3VdbeSorterWrite(db, pC, nKey); - if( rc==SQLITE_OK ){ + if( isSorter(pC) ){ + rc = sqlite3VdbeSorterWrite(db, pC, pIn2); + }else{ + nKey = pIn2->n; + zKey = pIn2->z; rc = sqlite3BtreeInsert(pCrsr, zKey, nKey, "", 0, 0, pOp->p3, ((pOp->p5 & OPFLAG_USESEEKRESULT) ? pC->seekResult : 0) - ); + ); assert( pC->deferredMoveto==0 ); + pC->cacheStatus = CACHE_STALE; } - pC->cacheStatus = CACHE_STALE; } } break; diff --git a/src/vdbeInt.h b/src/vdbeInt.h index 70d80c9d38..68ad8dcbf7 100644 --- a/src/vdbeInt.h +++ b/src/vdbeInt.h @@ -403,13 +403,15 @@ void sqlite3VdbeMemStoreType(Mem *pMem); # define sqlite3VdbeSorterRowkey(Y,Z) SQLITE_OK # define sqlite3VdbeSorterRewind(X,Y,Z) SQLITE_OK # define sqlite3VdbeSorterNext(X,Y,Z) SQLITE_OK +# define sqlite3VdbeSorterCompare(X,Y,Z) SQLITE_OK #else int sqlite3VdbeSorterInit(sqlite3 *, VdbeCursor *); -int sqlite3VdbeSorterWrite(sqlite3 *, VdbeCursor *, int); void sqlite3VdbeSorterClose(sqlite3 *, VdbeCursor *); int sqlite3VdbeSorterRowkey(VdbeCursor *, Mem *); -int sqlite3VdbeSorterRewind(sqlite3 *, VdbeCursor *, int *); int sqlite3VdbeSorterNext(sqlite3 *, VdbeCursor *, int *); +int sqlite3VdbeSorterRewind(sqlite3 *, VdbeCursor *, int *); +int sqlite3VdbeSorterWrite(sqlite3 *, VdbeCursor *, Mem *); +int sqlite3VdbeSorterCompare(VdbeCursor *, Mem *, int *); #endif #if !defined(SQLITE_OMIT_SHARED_CACHE) && SQLITE_THREADSAFE>0 diff --git a/src/vdbesort.c b/src/vdbesort.c index c3214b3afe..3ab5d7930a 100644 --- a/src/vdbesort.c +++ b/src/vdbesort.c @@ -21,6 +21,7 @@ #ifndef SQLITE_OMIT_MERGE_SORT typedef struct VdbeSorterIter VdbeSorterIter; +typedef struct SorterRecord SorterRecord; /* ** NOTES ON DATA STRUCTURE USED FOR N-WAY MERGES: @@ -92,8 +93,7 @@ typedef struct VdbeSorterIter VdbeSorterIter; ** being merged (rounded up to the next power of 2). */ struct VdbeSorter { - int nWorking; /* Start a new b-tree after this many pages */ - int nBtree; /* Current size of b-tree contents as PMA */ + int nInMemory; /* Current size of b-tree contents as PMA */ int nTree; /* Used size of aTree/aIter (power of 2) */ VdbeSorterIter *aIter; /* Array of iterators to merge */ int *aTree; /* Current state of incremental merge */ @@ -101,6 +101,9 @@ struct VdbeSorter { i64 iReadOff; /* Current read offset within file pTemp1 */ sqlite3_file *pTemp1; /* PMA file 1 */ int nPMA; /* Number of PMAs stored in pTemp1 */ + SorterRecord *pRecord; /* Head of in-memory record list */ + int nLimit1; /* Minimum PMA size, in bytes */ + int nLimit2; /* Maximum PMA size, in bytes */ }; /* @@ -117,6 +120,17 @@ struct VdbeSorterIter { u8 *aKey; /* Pointer to current key */ }; +/* +** A structure to store a single record. All in-memory records are connected +** together into a linked list headed at VdbeSorter.pRecord using the +** SorterRecord.pNext pointer. +*/ +struct SorterRecord { + void *pVal; + int nVal; + SorterRecord *pNext; +}; + /* Minimum allowable value for the VdbeSorter.nWorking variable */ #define SORTER_MIN_WORKING 10 @@ -275,6 +289,50 @@ static int vdbeSorterIterInit( return rc; } + +/* +** Compare key1 (buffer pKey1, size nKey1 bytes) with key2 (buffer pKey2, +** size nKey2 bytes). Argument pKeyInfo supplies the collation functions +** used by the comparison. If an error occurs, return an SQLite error code. +** Otherwise, return SQLITE_OK and set *pRes to a negative, zero or positive +** value, depending on whether key1 is smaller, equal to or larger than key2. +** +** If the bOmitRowid argument is non-zero, assume both keys end in a rowid +** field. For the purposes of the comparison, ignore it. Also, if bOmitRowid +** is true and key1 contains even a single NULL value, it is considered to +** be less than key2. Even if key2 also contains NULL values. +*/ +static int vdbeSorterCompare( + KeyInfo *pKeyInfo, /* Collation functions to use in comparison */ + int bOmitRowid, /* Ignore rowid field at end of keys */ + void *pKey1, int nKey1, /* Left side of comparison */ + void *pKey2, int nKey2, /* Right side of comparison */ + int *pRes /* OUT: Result of comparison */ +){ + char aSpace[150]; + UnpackedRecord *r2; + int i; + + r2 = sqlite3VdbeRecordUnpack(pKeyInfo, nKey2, pKey2, aSpace, sizeof(aSpace)); + if( r2==0 ) return SQLITE_NOMEM; + if( bOmitRowid ){ + for(i=0; inField-1; i++){ + if( r2->aMem[i].flags & MEM_Null ){ + *pRes = -1; + sqlite3VdbeDeleteUnpackedRecord(r2); + return SQLITE_OK; + } + } + r2->flags |= UNPACKED_PREFIX_MATCH; + r2->nField--; + assert( r2->nField>0 ); + } + + *pRes = sqlite3VdbeRecordCompare(nKey1, pKey1, r2); + sqlite3VdbeDeleteUnpackedRecord(r2); + return SQLITE_OK; +} + /* ** This function is called to compare two iterator keys when merging ** multiple b-tree segments. Parameter iOut is the index of the aTree[] @@ -306,20 +364,16 @@ static int vdbeSorterDoCompare(VdbeCursor *pCsr, int iOut){ }else if( p2->pFile==0 ){ iRes = i1; }else{ - char aSpace[150]; - UnpackedRecord *r1; - - r1 = sqlite3VdbeRecordUnpack( - pCsr->pKeyInfo, p1->nKey, p1->aKey, aSpace, sizeof(aSpace) + int res; + int rc = vdbeSorterCompare( + pCsr->pKeyInfo, 0, p1->aKey, p1->nKey, p2->aKey, p2->nKey, &res ); - if( r1==0 ) return SQLITE_NOMEM; - - if( sqlite3VdbeRecordCompare(p2->nKey, p2->aKey, r1)>=0 ){ + if( rc!=SQLITE_OK ) return rc; + if( res<=0 ){ iRes = i1; }else{ iRes = i2; } - sqlite3VdbeDeleteUnpackedRecord(r1); } pSorter->aTree[iOut] = iRes; @@ -330,9 +384,32 @@ static int vdbeSorterDoCompare(VdbeCursor *pCsr, int iOut){ ** Initialize the temporary index cursor just opened as a sorter cursor. */ int sqlite3VdbeSorterInit(sqlite3 *db, VdbeCursor *pCsr){ - assert( pCsr->pKeyInfo && pCsr->pBt ); + int pgsz; /* Page size of main database */ + + assert( pCsr->pKeyInfo && pCsr->pBt==0 ); pCsr->pSorter = sqlite3DbMallocZero(db, sizeof(VdbeSorter)); - return (pCsr->pSorter ? SQLITE_OK : SQLITE_NOMEM); + if( pCsr->pSorter==0 ){ + return SQLITE_NOMEM; + } + + pgsz = sqlite3BtreeGetPageSize(db->aDb[0].pBt); + pCsr->pSorter->nLimit1 = 10 * pgsz; + pCsr->pSorter->nLimit2 = db->aDb[0].pSchema->cache_size * pgsz; + + return SQLITE_OK; +} + +/* +** Free the list of sorted records starting at pRecord. +*/ +static void vdbeSorterRecordFree(sqlite3 *db, SorterRecord *pRecord){ + SorterRecord *p; + SorterRecord *pNext; + for(p=pRecord; p; p=pNext){ + pNext = p->pNext; + sqlite3DbFree(db, p->pVal); + sqlite3DbFree(db, p); + } } /* @@ -351,6 +428,7 @@ void sqlite3VdbeSorterClose(sqlite3 *db, VdbeCursor *pCsr){ if( pSorter->pTemp1 ){ sqlite3OsCloseFree(pSorter->pTemp1); } + vdbeSorterRecordFree(db, pSorter->pRecord); sqlite3DbFree(db, pSorter); pCsr->pSorter = 0; } @@ -370,10 +448,128 @@ static int vdbeSorterOpenTempFile(sqlite3 *db, sqlite3_file **ppFile){ ); } +/* +** Attemp to merge the two sorted lists p1 and p2 into a single list. If no +** error occurs set *ppOut to the head of the new list and return SQLITE_OK. +*/ +static int vdbeSorterMerge( + sqlite3 *db, /* Database handle */ + KeyInfo *pKeyInfo, /* Collation functions to use in comparison */ + SorterRecord *p1, /* First list to merge */ + SorterRecord *p2, /* Second list to merge */ + SorterRecord **ppOut /* OUT: Head of merged list */ +){ + int rc = SQLITE_OK; + SorterRecord *pFinal = 0; + SorterRecord **pp = &pFinal; + + while( p1 || p2 ){ + if( p1==0 ){ + *pp = p2; + p2 = 0; + }else if( p2==0 ){ + *pp = p1; + p1 = 0; + }else{ + int res; + rc = vdbeSorterCompare( + pKeyInfo, 0, p1->pVal, p1->nVal, p2->pVal, p2->nVal, &res + ); + if( rc!=SQLITE_OK ){ + vdbeSorterRecordFree(db, p1); + vdbeSorterRecordFree(db, p2); + vdbeSorterRecordFree(db, pFinal); + pFinal = 0; + break; + } + if( res<=0 ){ + *pp = p1; + pp = &p1->pNext; + p1 = p1->pNext; + }else{ + *pp = p2; + pp = &p2->pNext; + p2 = p2->pNext; + } + *pp = 0; + } + } + + *ppOut = pFinal; + return rc; +} /* -** Write the current contents of the b-tree to a PMA. Return SQLITE_OK -** if successful, or an SQLite error code otherwise. +** Sort the linked list of records headed at pCsr->pRecord. Return SQLITE_OK +** if successful, or an SQLite error code (i.e. SQLITE_NOMEM) if an error +** occurs. +*/ +static int vdbeSorterSort(sqlite3 *db, VdbeCursor *pCsr){ + int rc = SQLITE_OK; + int i; + SorterRecord **aSlot; + SorterRecord *p; + VdbeSorter *pSorter = pCsr->pSorter; + KeyInfo *pKeyInfo = pCsr->pKeyInfo; + + aSlot = (SorterRecord **)sqlite3MallocZero(64 * sizeof(SorterRecord *)); + if( !aSlot ){ + return SQLITE_NOMEM; + } + + p = pSorter->pRecord; + while( p ){ + SorterRecord *pNext = p->pNext; + p->pNext = 0; + for(i=0; rc==SQLITE_OK && aSlot[i]; i++){ + rc = vdbeSorterMerge(db, pKeyInfo, p, aSlot[i], &p); + aSlot[i] = 0; + } + if( rc!=SQLITE_OK ){ + vdbeSorterRecordFree(db, pNext); + break; + } + aSlot[i] = p; + p = pNext; + } + + p = 0; + for(i=0; i<64; i++){ + if( rc==SQLITE_OK ){ + rc = vdbeSorterMerge(db, pKeyInfo, p, aSlot[i], &p); + }else{ + vdbeSorterRecordFree(db, aSlot[i]); + } + } + pSorter->pRecord = p; + +#if 0 + { + SorterRecord *pTmp1 = 0; + SorterRecord *pTmp2; + for(pTmp2=pSorter->pRecord; pTmp2 && rc==SQLITE_OK; pTmp2=pTmp2->pNext){ + if( pTmp1 ){ + int res; + rc = vdbeSorterCompare(pKeyInfo, + 0, pTmp1->pVal, pTmp1->nVal, pTmp2->pVal, pTmp2->nVal, &res + ); + assert( rc!=SQLITE_OK || res<0 ); + } + pTmp1 = pTmp2; + } + } +#endif + + if( rc!=SQLITE_OK ){ + } + sqlite3_free(aSlot); + return rc; +} + + +/* +** Write the current contents of the in-memory linked-list to a PMA. Return +** SQLITE_OK if successful, or an SQLite error code otherwise. ** ** The format of a PMA is: ** @@ -384,19 +580,19 @@ static int vdbeSorterOpenTempFile(sqlite3 *db, sqlite3_file **ppFile){ ** Each record consists of a varint followed by a blob of data (the ** key). The varint is the number of bytes in the blob of data. */ -static int vdbeSorterBtreeToPMA(sqlite3 *db, VdbeCursor *pCsr){ +static int vdbeSorterListToPMA(sqlite3 *db, VdbeCursor *pCsr){ int rc = SQLITE_OK; /* Return code */ VdbeSorter *pSorter = pCsr->pSorter; - int res = 0; - /* sqlite3BtreeFirst() cannot fail because sorter btrees are always held - ** in memory and so an I/O error is not possible. */ - rc = sqlite3BtreeFirst(pCsr->pCursor, &res); - if( NEVER(rc!=SQLITE_OK) || res ) return rc; - assert( pSorter->nBtree>0 ); + if( pSorter->nInMemory==0 ){ + assert( pSorter->pRecord==0 ); + return rc; + } + + rc = vdbeSorterSort(db, pCsr); /* If the first temporary PMA file has not been opened, open it now. */ - if( pSorter->pTemp1==0 ){ + if( rc==SQLITE_OK && pSorter->pTemp1==0 ){ rc = vdbeSorterOpenTempFile(db, &pSorter->pTemp1); assert( rc!=SQLITE_OK || pSorter->pTemp1 ); assert( pSorter->iWriteOff==0 ); @@ -404,129 +600,89 @@ static int vdbeSorterBtreeToPMA(sqlite3 *db, VdbeCursor *pCsr){ } if( rc==SQLITE_OK ){ - i64 iWriteOff = pSorter->iWriteOff; - void *aMalloc = 0; /* Array used to hold a single record */ - int nMalloc = 0; /* Allocated size of aMalloc[] in bytes */ + i64 iOff = pSorter->iWriteOff; + SorterRecord *p; + SorterRecord *pNext = 0; pSorter->nPMA++; - for( - rc = vdbeSorterWriteVarint(pSorter->pTemp1, pSorter->nBtree, &iWriteOff); - rc==SQLITE_OK && res==0; - rc = sqlite3BtreeNext(pCsr->pCursor, &res) - ){ - i64 nKey; /* Size of this key in bytes */ + rc = vdbeSorterWriteVarint(pSorter->pTemp1, pSorter->nInMemory, &iOff); + for(p=pSorter->pRecord; rc==SQLITE_OK && p; p=pNext){ + pNext = p->pNext; + rc = vdbeSorterWriteVarint(pSorter->pTemp1, p->nVal, &iOff); - /* Write the size of the record in bytes to the output file */ - (void)sqlite3BtreeKeySize(pCsr->pCursor, &nKey); - rc = vdbeSorterWriteVarint(pSorter->pTemp1, nKey, &iWriteOff); - - /* Make sure the aMalloc[] buffer is large enough for the record */ - if( rc==SQLITE_OK && nKey>nMalloc ){ - aMalloc = sqlite3DbReallocOrFree(db, aMalloc, nKey); - if( !aMalloc ){ - rc = SQLITE_NOMEM; - }else{ - nMalloc = nKey; - } - } - - /* Write the record itself to the output file */ if( rc==SQLITE_OK ){ - /* sqlite3BtreeKey() cannot fail because sorter btrees held in memory */ - rc = sqlite3BtreeKey(pCsr->pCursor, 0, nKey, aMalloc); - if( ALWAYS(rc==SQLITE_OK) ){ - rc = sqlite3OsWrite(pSorter->pTemp1, aMalloc, nKey, iWriteOff); - iWriteOff += nKey; - } + rc = sqlite3OsWrite(pSorter->pTemp1, p->pVal, p->nVal, iOff); + iOff += p->nVal; } - if( rc!=SQLITE_OK ) break; + sqlite3DbFree(db, p->pVal); + sqlite3DbFree(db, p); } /* This assert verifies that unless an error has occurred, the size of ** the PMA on disk is the same as the expected size stored in - ** pSorter->nBtree. */ - assert( rc!=SQLITE_OK || pSorter->nBtree==( - iWriteOff-pSorter->iWriteOff-sqlite3VarintLen(pSorter->nBtree) + ** pSorter->nInMemory. */ + assert( rc!=SQLITE_OK || pSorter->nInMemory==( + iOff-pSorter->iWriteOff-sqlite3VarintLen(pSorter->nInMemory) )); - pSorter->iWriteOff = iWriteOff; - sqlite3DbFree(db, aMalloc); + pSorter->iWriteOff = iOff; + pSorter->pRecord = p; } - pSorter->nBtree = 0; return rc; } /* -** This function is called on a sorter cursor by the VDBE before each row -** is inserted into VdbeCursor.pCsr. Argument nKey is the size of the key, in -** bytes, about to be inserted. -** -** If it is determined that the temporary b-tree accessed via VdbeCursor.pCsr -** is large enough, its contents are written to a sorted PMA on disk and the -** tree emptied. This prevents the b-tree (which must be small enough to -** fit entirely in the cache in order to support efficient inserts) from -** growing too large. -** -** An SQLite error code is returned if an error occurs. Otherwise, SQLITE_OK. +** Add a record to the sorter. */ -int sqlite3VdbeSorterWrite(sqlite3 *db, VdbeCursor *pCsr, int nKey){ - int rc = SQLITE_OK; /* Return code */ +int sqlite3VdbeSorterWrite( + sqlite3 *db, /* Database handle */ + VdbeCursor *pCsr, /* Sorter cursor */ + Mem *pVal /* Memory cell containing record */ +){ VdbeSorter *pSorter = pCsr->pSorter; - if( pSorter ){ - Pager *pPager = sqlite3BtreePager(pCsr->pBt); - int nPage; /* Current size of temporary file in pages */ + int rc; + SorterRecord *pNew; - /* Sorters never spill to disk */ - assert( sqlite3PagerFile(pPager)->pMethods==0 ); + assert( pSorter ); + pSorter->nInMemory += sqlite3VarintLen(pVal->n) + pVal->n; - /* Determine how many pages the temporary b-tree has grown to */ - sqlite3PagerPagecount(pPager, &nPage); - - /* If pSorter->nWorking is still zero, but the temporary file has been - ** created in the file-system, then the most recent insert into the - ** current b-tree segment probably caused the cache to overflow (it is - ** also possible that sqlite3_release_memory() was called). So set the - ** size of the working set to a little less than the current size of the - ** file in pages. */ - if( pSorter->nWorking==0 && sqlite3PagerUnderStress(pPager) ){ - pSorter->nWorking = nPage-5; - if( pSorter->nWorkingnWorking = SORTER_MIN_WORKING; - } + pNew = (SorterRecord *)sqlite3DbMallocZero(db, sizeof(SorterRecord)); + if( pNew==0 ){ + rc = SQLITE_NOMEM; + }else{ + rc = sqlite3VdbeMemMakeWriteable(pVal); + if( rc==SQLITE_OK ){ + pNew->pVal = pVal->z; + pNew->nVal = pVal->n; + pVal->zMalloc = 0; + sqlite3VdbeMemSetNull(pVal); + pNew->pNext = pSorter->pRecord; + pSorter->pRecord = pNew; + }else{ + sqlite3DbFree(db, pNew); + rc = SQLITE_NOMEM; } - - /* If the number of pages used by the current b-tree segment is greater - ** than the size of the working set (VdbeSorter.nWorking), start a new - ** segment b-tree. */ - if( pSorter->nWorking && nPage>=pSorter->nWorking ){ - BtCursor *p = pCsr->pCursor;/* Cursor structure to close and reopen */ - int iRoot; /* Root page of new tree */ - - /* Copy the current contents of the b-tree into a PMA in sorted order. - ** Close the currently open b-tree cursor. */ - rc = vdbeSorterBtreeToPMA(db, pCsr); - sqlite3BtreeCloseCursor(p); - - if( rc==SQLITE_OK ){ - rc = sqlite3BtreeDropTable(pCsr->pBt, 2, 0); -#ifdef SQLITE_DEBUG - sqlite3PagerPagecount(pPager, &nPage); - assert( rc!=SQLITE_OK || nPage==1 ); -#endif - } - if( rc==SQLITE_OK ){ - rc = sqlite3BtreeCreateTable(pCsr->pBt, &iRoot, BTREE_BLOBKEY); - } - if( rc==SQLITE_OK ){ - assert( iRoot==2 ); - rc = sqlite3BtreeCursor(pCsr->pBt, iRoot, 1, pCsr->pKeyInfo, p); - } - } - - pSorter->nBtree += sqlite3VarintLen(nKey) + nKey; } + + /* See if the contents of the sorter should now be written out. They + ** are written out when either of the following are true: + ** + ** * The total memory allocated for the in-memory list is greater + ** than (page-size * cache-size), or + ** + ** * The total memory allocated for the in-memory list is greater + ** than (page-size * 10) and sqlite3HeapNearlyFull() returns true. + */ + if( rc==SQLITE_OK && ( + (pSorter->nInMemory>pSorter->nLimit2) + || (pSorter->nInMemory>pSorter->nLimit1 && sqlite3HeapNearlyFull()) + )){ + rc = vdbeSorterListToPMA(db, pCsr); + pSorter->nInMemory = 0; + } + return rc; } @@ -577,8 +733,7 @@ int sqlite3VdbeSorterRewind(sqlite3 *db, VdbeCursor *pCsr, int *pbEof){ assert( pSorter ); /* Write the current b-tree to a PMA. Close the b-tree cursor. */ - rc = vdbeSorterBtreeToPMA(db, pCsr); - sqlite3BtreeCloseCursor(pCsr->pCursor); + rc = vdbeSorterListToPMA(db, pCsr); if( rc!=SQLITE_OK ) return rc; if( pSorter->nPMA==0 ){ *pbEof = 1; @@ -692,14 +847,7 @@ int sqlite3VdbeSorterRowkey(VdbeCursor *pCsr, Mem *pOut){ VdbeSorterIter *pIter; pIter = &pSorter->aIter[ pSorter->aTree[1] ]; - - /* Coverage testing note: As things are currently, this call will always - ** succeed. This is because the memory cell passed by the VDBE layer - ** happens to be the same one as was used to assemble the keys before they - ** were passed to the sorter - meaning it is always large enough for the - ** largest key. But this could change very easily, so we leave the call - ** to sqlite3VdbeMemGrow() in. */ - if( NEVER(sqlite3VdbeMemGrow(pOut, pIter->nKey, 0)) ){ + if( sqlite3VdbeMemGrow(pOut, pIter->nKey, 0) ){ return SQLITE_NOMEM; } pOut->n = pIter->nKey; @@ -709,4 +857,30 @@ int sqlite3VdbeSorterRowkey(VdbeCursor *pCsr, Mem *pOut){ return SQLITE_OK; } +/* +** Compare the key in memory cell pVal with the key that the sorter cursor +** passed as the first argument currently points to. For the purposes of +** the comparison, ignore the rowid field at the end of each record. +** +** If an error occurs, return an SQLite error code (i.e. SQLITE_NOMEM). +** Otherwise, set *pRes to a negative, zero or positive value if the +** key in pVal is smaller than, equal to or larger than the current sorter +** key. +*/ +int sqlite3VdbeSorterCompare( + VdbeCursor *pCsr, /* Sorter cursor */ + Mem *pVal, /* Value to compare to current sorter key */ + int *pRes /* OUT: Result of comparison */ +){ + int rc; + VdbeSorter *pSorter = pCsr->pSorter; + VdbeSorterIter *pIter; + pIter = &pSorter->aIter[ pSorter->aTree[1] ]; + rc = vdbeSorterCompare(pCsr->pKeyInfo, 1, + pVal->z, pVal->n, pIter->aKey, pIter->nKey, pRes + ); + assert( rc!=SQLITE_OK || *pRes<=0 ); + return rc; +} + #endif /* #ifndef SQLITE_OMIT_MERGE_SORT */ diff --git a/test/index4.test b/test/index4.test index 6400e34e4e..018ed744a1 100644 --- a/test/index4.test +++ b/test/index4.test @@ -108,5 +108,19 @@ do_execsql_test 1.8 { PRAGMA integrity_check } {ok} +do_execsql_test 2.1 { + BEGIN; + CREATE TABLE t2(x); + INSERT INTO t2 VALUES(14); + INSERT INTO t2 VALUES(35); + INSERT INTO t2 VALUES(15); + INSERT INTO t2 VALUES(35); + INSERT INTO t2 VALUES(16); + COMMIT; +} +do_catchsql_test 2.2 { + CREATE UNIQUE INDEX i3 ON t2(x); +} {1 {indexed columns are not unique}} + finish_test