The performance increase in the previous check-in of this branch was due to

the revised loop in initMemArray() and reordering fields of Mem - not the
call the memcpy().  Changing the code to avoid memcpy() results in an even
better gain, and code that is far less dodgy.

FossilOrigin-Name: d74aa979530d4236f5900d2ef998b27065d352d7c18bcd822e5c8f1041a1a81c

diff --git a/manifest b/manifest
index d712c505fdd7564dc64a0f430c8041994bb83207..a4673e8d7efe1e0347aec9aa881670e163c9b291 100644 (file)
--- a/manifest
+++ b/manifest
@@ -1,5 +1,5 @@
-C An\soptimization\sto\sinitMemArray()\ssaves\salmost\s500K\scycles.\s\sBut\sit\sseems\sa\nlittle\sdodgy.\s\sI\swant\sto\sthink\sabout\sthis\smore\sbefore\smerging\sto\strunk.\nPerhaps\sthere\sis\sa\scleaner\sway\sto\saccomplish\sthe\ssame.
-D 2022-02-28T03:25:13.694
+C The\sperformance\sincrease\sin\sthe\sprevious\scheck-in\sof\sthis\sbranch\swas\sdue\sto\nthe\srevised\sloop\sin\sinitMemArray()\sand\sreordering\sfields\sof\sMem\s-\snot\sthe\ncall\sthe\smemcpy().\s\sChanging\sthe\scode\sto\savoid\smemcpy()\sresults\sin\san\seven\nbetter\sgain,\sand\scode\sthat\sis\sfar\sless\sdodgy.
+D 2022-02-28T12:08:09.462
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
 F LICENSE.md df5091916dbb40e6e9686186587125e1b2ff51f022cc334e886c19a0e9982724
@@ -626,9 +626,9 @@ F src/util.c 602fe229f32a96ceccae4f40824129669582096f7c355f53dbac156c9fecef23
 F src/vacuum.c 6c38ddc52f0619865c91dae9c441d4d48bf3040d7dc1bc5b22da1e45547ed0b3
 F src/vdbe.c 4b969ebe6b61f87a90aebf817bc6ebda5075fe56987591091a9bf22556262484
 F src/vdbe.h a1d0e3b934e835e73edd146f2e7c4eadb711b5c9875c18159a57483fd78e550e
-F src/vdbeInt.h 1336e8e23cec21df0dd83d61d0981e3ce72dd19923e9792182066affa2adbf73
-F src/vdbeapi.c 9daec6f3382a150f1da47fc74f2b3aa5b1790bd9dff322539887d25f18054ca9
-F src/vdbeaux.c e5dfd98fe562475d6519af81667b99d237f4999c4709be219d1084ea1fb0af1f
+F src/vdbeInt.h 2ff02995d153d30d362c99d27248b8a3bf825617a8e0d4d8d1e0231eca3bc4f7
+F src/vdbeapi.c 1c80efbe51118bbecc7279023e75d18edcfa4b3dc441287e1718ee70ad594f58
+F src/vdbeaux.c 0f3ce77dabb1f897cc7d3812ed162ec5d31d2298d1aa41b606f8048a742c4b9e
 F src/vdbeblob.c 5e61ce31aca17db8fb60395407457a8c1c7fb471dde405e0cd675974611dcfcd
 F src/vdbemem.c 7737f0b1c480a32b057849c804d2f21d5389649bb8be80f77ad75df700adc9a1
 F src/vdbesort.c 43756031ca7430f7aec3ef904824a7883c4ede783e51f280d99b9b65c0796e35
@@ -1944,11 +1944,8 @@ F vsixtest/vsixtest.tcl 6a9a6ab600c25a91a7acc6293828957a386a8a93
 F vsixtest/vsixtest.vcxproj.data 2ed517e100c66dc455b492e1a33350c1b20fbcdc
 F vsixtest/vsixtest.vcxproj.filters 37e51ffedcdb064aad6ff33b6148725226cd608e
 F vsixtest/vsixtest_TemporaryKey.pfx e5b1b036facdb453873e7084e1cae9102ccc67a0
-P bb520293d8c11518ba153b986662f081ebfd781d38eb624c509605fa9148f6e9
-R f523ab929bd4578ff1ab123801eb9b63
-T *branch * optimize-init-mem
-T *sym-optimize-init-mem *
-T -sym-trunk *
+P 7fefd8676110a53e6c98a697e2dbf820740fe602a1e83b6caa8d099c41a15d80
+R 0cbe8ba68087a6c2d38d406f08a0fa0f
 U drh
-Z 9ff8d5adc5911dd21d68c30eb1e14ee4
+Z d7015a158f91cecfa6d5c894c1220cae
 # Remove this line to create a well-formed Fossil manifest.

diff --git a/manifest.uuid b/manifest.uuid
index d40615c913224c8d6f5c6e288fcb2b51c186a149..7529ec2db4bed258dea013e1cd1a9bc9c4e27df9 100644 (file)
--- a/manifest.uuid
+++ b/manifest.uuid
@@ -1 +1 @@
-7fefd8676110a53e6c98a697e2dbf820740fe602a1e83b6caa8d099c41a15d80
\ No newline at end of file
+d74aa979530d4236f5900d2ef998b27065d352d7c18bcd822e5c8f1041a1a81c
\ No newline at end of file

diff --git a/src/vdbeInt.h b/src/vdbeInt.h
index dc284fb0bbb8640149297071937e560c1c5ae110..657c6f8cf6a9b0d5ae77d0d5b0377dc940c059f1 100644 (file)
--- a/src/vdbeInt.h
+++ b/src/vdbeInt.h
@@ -199,10 +199,6 @@ struct VdbeFrame {
 ** Internally, the vdbe manipulates nearly all SQL values as Mem
 ** structures. Each Mem struct may cache multiple representations (string,
 ** integer etc.) of the same value.
-**
-** Code uses offsetof() on this object.  Order of the fields is important.
-** Search for tag-20220228a to find all places that need to change when the
-** field order changes.
 */
 struct sqlite3_value {
   union MemValue {

diff --git a/src/vdbeapi.c b/src/vdbeapi.c
index 69f3d7b120e20d1b00424f3222c64a666a8f4c4c..83f7b0fd65fca0f57be41c72eff7b62fac9a1d24 100644 (file)
--- a/src/vdbeapi.c
+++ b/src/vdbeapi.c
@@ -1087,8 +1087,6 @@ int sqlite3_data_count(sqlite3_stmt *pStmt){
 
 /*
 ** Return a pointer to static memory containing an SQL NULL value.
-**
-** Must be revised if column order for Mem changes.  tag-20220228a.
 */
 static const Mem *columnNullValue(void){
   /* Even though the Mem structure contains an element

diff --git a/src/vdbeaux.c b/src/vdbeaux.c
index 0d91124c9dd01678e2d412021c3931aa65d407a3..54332caa3731132be4059d4550cd8db921a059f4 100644 (file)
--- a/src/vdbeaux.c
+++ b/src/vdbeaux.c
@@ -1839,28 +1839,24 @@ void sqlite3VdbePrintOp(FILE *pOut, int pc, VdbeOp *pOp){
 /*
 ** Initialize an array of N Mem element.
 **
-** This is a high-runner, so it is optimized by taking advantage of the
-** order of the fields in a Mem object and using memcpy() rather than
-** individually setting each field.  For each Mem, we need to set:
+** This is a high-runner, so only those fields that really do need to
+** be initialized are set.  The Mem structure is organized so that
+** the fields that get initialized are nearby and hopefully on the same
+** cache line.
 **
 **    Mem.flags = flags
 **    Mem.db = db
 **    Mem.szMalloc = 0
 **
 ** All other fields of Mem can safely remain uninitialized for now.  They
-** will be initialized before use.  The fields that are initialized by this
-** routine are grouped together so that they can be set using memcpy().
-**
-** tag-20220228a
+** will be initialized before use.
 */
 static void initMemArray(Mem *p, int N, sqlite3 *db, u16 flags){
   if( N>0 ){
-    Mem x;
-    x.flags = flags;
-    x.db = db;
-    x.szMalloc = 0;
     do{
-      memcpy(&p->flags, &x.flags, offsetof(Mem,uTemp)-offsetof(Mem,flags));
+      p->flags = flags;
+      p->db = db;
+      p->szMalloc = 0;
 #ifdef SQLITE_DEBUG
       p->pScopyFrom = 0;
 #endif