aboutsummaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
authordrh <>2023-04-08 13:01:34 +0000
committerdrh <>2023-04-08 13:01:34 +0000
commit2da07d9025f24ee10a6e2432394d36a291c96ea8 (patch)
tree536383e921068e744ab6433f3c712cf6008fb196 /src
parent6d47af64e4d59951c8e8dd0d064fd6c2e4fe13d8 (diff)
parent65aae44b87551bac15ff2a58ca7d7d7f86ea1114 (diff)
downloadsqlite-2da07d9025f24ee10a6e2432394d36a291c96ea8.tar.gz
sqlite-2da07d9025f24ee10a6e2432394d36a291c96ea8.zip
Optimizations to btree.c save about 4.5 million CPU cycles:
(1) Clone insertCell() into a separate insertCellFast() routine for use by sqlite3BtreeInsert(). (2) Mark allocateSpace() as always-inline. (3) Improved coalesence of adjacent free blocks in pageFreeArray(). FossilOrigin-Name: 5c12c400fe8eb4e86e14c69a6c34d0d78d9861e5d40a36c6a596a81c6dd65977
Diffstat (limited to 'src')
-rw-r--r--src/btree.c181
-rw-r--r--src/sqliteInt.h9
2 files changed, 146 insertions, 44 deletions
diff --git a/src/btree.c b/src/btree.c
index 2f3063c3a..0a3d71317 100644
--- a/src/btree.c
+++ b/src/btree.c
@@ -1746,7 +1746,7 @@ static u8 *pageFindSlot(MemPage *pPg, int nByte, int *pRc){
** allocation is being made in order to insert a new cell, so we will
** also end up needing a new cell pointer.
*/
-static int allocateSpace(MemPage *pPage, int nByte, int *pIdx){
+static SQLITE_INLINE int allocateSpace(MemPage *pPage, int nByte, int *pIdx){
const int hdr = pPage->hdrOffset; /* Local cache of pPage->hdrOffset */
u8 * const data = pPage->aData; /* Local cache of pPage->aData */
int top; /* First byte of cell content area */
@@ -1862,7 +1862,7 @@ static int freeSpace(MemPage *pPage, u16 iStart, u16 iSize){
assert( CORRUPT_DB || iEnd <= pPage->pBt->usableSize );
assert( sqlite3_mutex_held(pPage->pBt->mutex) );
assert( iSize>=4 ); /* Minimum cell size is 4 */
- assert( iStart<=pPage->pBt->usableSize-4 );
+ assert( CORRUPT_DB || iStart<=pPage->pBt->usableSize-4 );
/* The list of freeblocks must be in ascending order. Find the
** spot on the list where iStart should be inserted.
@@ -1919,6 +1919,11 @@ static int freeSpace(MemPage *pPage, u16 iStart, u16 iSize){
}
pTmp = &data[hdr+5];
x = get2byte(pTmp);
+ if( pPage->pBt->btsFlags & BTS_FAST_SECURE ){
+ /* Overwrite deleted information with zeros when the secure_delete
+ ** option is enabled */
+ memset(&data[iStart], 0, iSize);
+ }
if( iStart<=x ){
/* The new freeblock is at the beginning of the cell content area,
** so just extend the cell content area rather than create another
@@ -1930,14 +1935,9 @@ static int freeSpace(MemPage *pPage, u16 iStart, u16 iSize){
}else{
/* Insert the new freeblock into the freelist */
put2byte(&data[iPtr], iStart);
+ put2byte(&data[iStart], iFreeBlk);
+ put2byte(&data[iStart+2], iSize);
}
- if( pPage->pBt->btsFlags & BTS_FAST_SECURE ){
- /* Overwrite deleted information with zeros when the secure_delete
- ** option is enabled */
- memset(&data[iStart], 0, iSize);
- }
- put2byte(&data[iStart], iFreeBlk);
- put2byte(&data[iStart+2], iSize);
pPage->nFree += iOrigSize;
return SQLITE_OK;
}
@@ -7096,6 +7096,14 @@ static void dropCell(MemPage *pPage, int idx, int sz, int *pRC){
** in pTemp or the original pCell) and also record its index.
** Allocating a new entry in pPage->aCell[] implies that
** pPage->nOverflow is incremented.
+**
+** The insertCellFast() routine below works exactly the same as
+** insertCell() except that it lacks the pTemp and iChild parameters
+** which are assumed zero. Other than that, the two routines are the
+** same.
+**
+** Fixes or enhancements to this routine should be reflected in
+** insertCellFast()!
*/
static int insertCell(
MemPage *pPage, /* Page into which we are copying */
@@ -7118,14 +7126,103 @@ static int insertCell(
assert( sqlite3_mutex_held(pPage->pBt->mutex) );
assert( sz==pPage->xCellSize(pPage, pCell) || CORRUPT_DB );
assert( pPage->nFree>=0 );
+ assert( iChild>0 );
if( pPage->nOverflow || sz+2>pPage->nFree ){
if( pTemp ){
memcpy(pTemp, pCell, sz);
pCell = pTemp;
}
- if( iChild ){
- put4byte(pCell, iChild);
+ put4byte(pCell, iChild);
+ j = pPage->nOverflow++;
+ /* Comparison against ArraySize-1 since we hold back one extra slot
+ ** as a contingency. In other words, never need more than 3 overflow
+ ** slots but 4 are allocated, just to be safe. */
+ assert( j < ArraySize(pPage->apOvfl)-1 );
+ pPage->apOvfl[j] = pCell;
+ pPage->aiOvfl[j] = (u16)i;
+
+ /* When multiple overflows occur, they are always sequential and in
+ ** sorted order. This invariants arise because multiple overflows can
+ ** only occur when inserting divider cells into the parent page during
+ ** balancing, and the dividers are adjacent and sorted.
+ */
+ assert( j==0 || pPage->aiOvfl[j-1]<(u16)i ); /* Overflows in sorted order */
+ assert( j==0 || i==pPage->aiOvfl[j-1]+1 ); /* Overflows are sequential */
+ }else{
+ int rc = sqlite3PagerWrite(pPage->pDbPage);
+ if( NEVER(rc!=SQLITE_OK) ){
+ return rc;
}
+ assert( sqlite3PagerIswriteable(pPage->pDbPage) );
+ data = pPage->aData;
+ assert( &data[pPage->cellOffset]==pPage->aCellIdx );
+ rc = allocateSpace(pPage, sz, &idx);
+ if( rc ){ return rc; }
+ /* The allocateSpace() routine guarantees the following properties
+ ** if it returns successfully */
+ assert( idx >= 0 );
+ assert( idx >= pPage->cellOffset+2*pPage->nCell+2 || CORRUPT_DB );
+ assert( idx+sz <= (int)pPage->pBt->usableSize );
+ pPage->nFree -= (u16)(2 + sz);
+ /* In a corrupt database where an entry in the cell index section of
+ ** a btree page has a value of 3 or less, the pCell value might point
+ ** as many as 4 bytes in front of the start of the aData buffer for
+ ** the source page. Make sure this does not cause problems by not
+ ** reading the first 4 bytes */
+ memcpy(&data[idx+4], pCell+4, sz-4);
+ put4byte(&data[idx], iChild);
+ pIns = pPage->aCellIdx + i*2;
+ memmove(pIns+2, pIns, 2*(pPage->nCell - i));
+ put2byte(pIns, idx);
+ pPage->nCell++;
+ /* increment the cell count */
+ if( (++data[pPage->hdrOffset+4])==0 ) data[pPage->hdrOffset+3]++;
+ assert( get2byte(&data[pPage->hdrOffset+3])==pPage->nCell || CORRUPT_DB );
+#ifndef SQLITE_OMIT_AUTOVACUUM
+ if( pPage->pBt->autoVacuum ){
+ int rc2 = SQLITE_OK;
+ /* The cell may contain a pointer to an overflow page. If so, write
+ ** the entry for the overflow page into the pointer map.
+ */
+ ptrmapPutOvflPtr(pPage, pPage, pCell, &rc2);
+ if( rc2 ) return rc2;
+ }
+#endif
+ }
+ return SQLITE_OK;
+}
+
+/*
+** This variant of insertCell() assumes that the pTemp and iChild
+** parameters are both zero. Use this variant in sqlite3BtreeInsert()
+** for performance improvement, and also so that this variant is only
+** called from that one place, and is thus inlined, and thus runs must
+** faster.
+**
+** Fixes or enhancements to this routine should be reflected into
+** the insertCell() routine.
+*/
+static int insertCellFast(
+ MemPage *pPage, /* Page into which we are copying */
+ int i, /* New cell becomes the i-th cell of the page */
+ u8 *pCell, /* Content of the new cell */
+ int sz /* Bytes of content in pCell */
+){
+ int idx = 0; /* Where to write new cell content in data[] */
+ int j; /* Loop counter */
+ u8 *data; /* The content of the whole page */
+ u8 *pIns; /* The point in pPage->aCellIdx[] where no cell inserted */
+
+ assert( i>=0 && i<=pPage->nCell+pPage->nOverflow );
+ assert( MX_CELL(pPage->pBt)<=10921 );
+ assert( pPage->nCell<=MX_CELL(pPage->pBt) || CORRUPT_DB );
+ assert( pPage->nOverflow<=ArraySize(pPage->apOvfl) );
+ assert( ArraySize(pPage->apOvfl)==ArraySize(pPage->aiOvfl) );
+ assert( sqlite3_mutex_held(pPage->pBt->mutex) );
+ assert( sz==pPage->xCellSize(pPage, pCell) || CORRUPT_DB );
+ assert( pPage->nFree>=0 );
+ assert( pPage->nOverflow==0 );
+ if( sz+2>pPage->nFree ){
j = pPage->nOverflow++;
/* Comparison against ArraySize-1 since we hold back one extra slot
** as a contingency. In other words, never need more than 3 overflow
@@ -7157,17 +7254,7 @@ static int insertCell(
assert( idx >= pPage->cellOffset+2*pPage->nCell+2 || CORRUPT_DB );
assert( idx+sz <= (int)pPage->pBt->usableSize );
pPage->nFree -= (u16)(2 + sz);
- if( iChild ){
- /* In a corrupt database where an entry in the cell index section of
- ** a btree page has a value of 3 or less, the pCell value might point
- ** as many as 4 bytes in front of the start of the aData buffer for
- ** the source page. Make sure this does not cause problems by not
- ** reading the first 4 bytes */
- memcpy(&data[idx+4], pCell+4, sz-4);
- put4byte(&data[idx], iChild);
- }else{
- memcpy(&data[idx], pCell, sz);
- }
+ memcpy(&data[idx], pCell, sz);
pIns = pPage->aCellIdx + i*2;
memmove(pIns+2, pIns, 2*(pPage->nCell - i));
put2byte(pIns, idx);
@@ -7496,42 +7583,50 @@ static int pageFreeArray(
u8 * const pEnd = &aData[pPg->pBt->usableSize];
u8 * const pStart = &aData[pPg->hdrOffset + 8 + pPg->childPtrSize];
int nRet = 0;
- int i;
+ int i, j;
int iEnd = iFirst + nCell;
- u8 *pFree = 0; /* \__ Parameters for pending call to */
- int szFree = 0; /* / freeSpace() */
+ int nFree = 0;
+ int aOfst[10];
+ int aAfter[10];
for(i=iFirst; i<iEnd; i++){
u8 *pCell = pCArray->apCell[i];
if( SQLITE_WITHIN(pCell, pStart, pEnd) ){
int sz;
+ int iAfter;
+ int iOfst;
/* No need to use cachedCellSize() here. The sizes of all cells that
** are to be freed have already been computing while deciding which
** cells need freeing */
sz = pCArray->szCell[i]; assert( sz>0 );
- if( pFree!=(pCell + sz) ){
- if( pFree ){
- assert( pFree>aData && (pFree - aData)<65536 );
- freeSpace(pPg, (u16)(pFree - aData), szFree);
+ iOfst = (u16)(pCell - aData);
+ iAfter = iOfst+sz;
+ for(j=0; j<nFree; j++){
+ if( aOfst[j]==iAfter ){
+ aOfst[j] = iOfst;
+ break;
+ }else if( aAfter[j]==iOfst ){
+ aAfter[j] = iAfter;
+ break;
}
- pFree = pCell;
- szFree = sz;
- if( pFree+sz>pEnd ){
- return 0;
+ }
+ if( j>=nFree ){
+ if( nFree>=sizeof(aOfst)/sizeof(aOfst[0]) ){
+ for(j=0; j<nFree; j++){
+ freeSpace(pPg, aOfst[j], aAfter[j]-aOfst[j]);
+ }
+ nFree = 0;
}
- }else{
- /* The current cell is adjacent to and before the pFree cell.
- ** Combine the two regions into one to reduce the number of calls
- ** to freeSpace(). */
- pFree = pCell;
- szFree += sz;
+ aOfst[nFree] = iOfst;
+ aAfter[nFree] = iAfter;
+ if( &aData[iAfter]>pEnd ) return 0;
+ nFree++;
}
nRet++;
}
}
- if( pFree ){
- assert( pFree>aData && (pFree - aData)<65536 );
- freeSpace(pPg, (u16)(pFree - aData), szFree);
+ for(j=0; j<nFree; j++){
+ freeSpace(pPg, aOfst[j], aAfter[j]-aOfst[j]);
}
return nRet;
}
@@ -9312,7 +9407,7 @@ int sqlite3BtreeInsert(
}else{
assert( pPage->leaf );
}
- rc = insertCell(pPage, idx, newCell, szNew, 0, 0);
+ rc = insertCellFast(pPage, idx, newCell, szNew);
assert( pPage->nOverflow==0 || rc==SQLITE_OK );
assert( rc!=SQLITE_OK || pPage->nCell>0 || pPage->nOverflow>0 );
diff --git a/src/sqliteInt.h b/src/sqliteInt.h
index 9b189124d..f33f6eb20 100644
--- a/src/sqliteInt.h
+++ b/src/sqliteInt.h
@@ -281,15 +281,22 @@
#endif
/*
-** A macro to hint to the compiler that a function should not be
+** Macros to hint to the compiler that a function should or should not be
** inlined.
*/
#if defined(__GNUC__)
# define SQLITE_NOINLINE __attribute__((noinline))
+# define SQLITE_INLINE __attribute__((always_inline)) inline
#elif defined(_MSC_VER) && _MSC_VER>=1310
# define SQLITE_NOINLINE __declspec(noinline)
+# define SQLITE_INLINE __forceinline
#else
# define SQLITE_NOINLINE
+# define SQLITE_INLINE
+#endif
+#if defined(SQLITE_COVERAGE_TEST)
+# undef SQLITE_INLINE
+# define SQLITE_INLINE
#endif
/*