Changeset: 63cc5dda9632 for MonetDB
URL: https://dev.monetdb.org/hg/MonetDB?cmd=changeset;node=63cc5dda9632
Modified Files:
monetdb5/optimizer/opt_pipes.c
monetdb5/optimizer/opt_prelude.c
monetdb5/optimizer/opt_prelude.h
monetdb5/optimizer/opt_projection2candidate.c
Branch: projection2candidate
Log Message:
first version of a "showcase" MAL optimize to eliminate algrebra.projection()
calls
by making subsequent MAL operations use the respective candidate list directly.
For now, all code is disabled by compiler directives ("#if 0 ... #endif")
because most of the code does not work, yet;
seemingly, the available MAL operations with candidates
might not yet work as assumed, or they get used wrongly.
Comments in the code detail the problems that occur with TPC-H.
diffs (truncated from 461 to 300 lines):
diff --git a/monetdb5/optimizer/opt_pipes.c b/monetdb5/optimizer/opt_pipes.c
--- a/monetdb5/optimizer/opt_pipes.c
+++ b/monetdb5/optimizer/opt_pipes.c
@@ -91,7 +91,10 @@ static struct PIPELINES {
"optimizer.candidates();"
"optimizer.postfix();"
"optimizer.deadcode();"
-// "optimizer.projection2candidate();" awaiting the new batcalc api
+ "optimizer.projection2candidate();"
+ "optimizer.deadcode();"
+ "optimizer.projectionpath();"
+ "optimizer.deadcode();"
// "optimizer.oltp();"awaiting the autocommit front-end changes
"optimizer.wlc();"
"optimizer.garbageCollector();",
@@ -128,7 +131,10 @@ static struct PIPELINES {
"optimizer.candidates();"
"optimizer.postfix();"
"optimizer.deadcode();"
-// "optimizer.projection2candidate();" awaiting the new batcalc api
+ "optimizer.projection2candidate();"
+ "optimizer.deadcode();"
+ "optimizer.projectionpath();"
+ "optimizer.deadcode();"
// "optimizer.oltp();"awaiting the autocommit front-end changes
"optimizer.wlc();"
"optimizer.garbageCollector();",
@@ -170,7 +176,10 @@ static struct PIPELINES {
"optimizer.candidates();"
"optimizer.postfix();"
"optimizer.deadcode();"
-// "optimizer.projection2candidate();" awaiting the new batcalc api
+ "optimizer.projection2candidate();"
+ "optimizer.deadcode();"
+ "optimizer.projectionpath();"
+ "optimizer.deadcode();"
// "optimizer.oltp();"awaiting the autocommit front-end changes
"optimizer.wlc();"
"optimizer.garbageCollector();",
@@ -199,7 +208,7 @@ static struct PIPELINES {
"optimizer.aliases();"
"optimizer.constants();"
"optimizer.commonTerms();"
- "optimizer.projectionpath();"
+ //"optimizer.projectionpath();"
"optimizer.reorder();"
"optimizer.deadcode();"
// "optimizer.reduce();" deprecated
@@ -211,7 +220,10 @@ static struct PIPELINES {
"optimizer.candidates();"
"optimizer.postfix();"
"optimizer.deadcode();"
-// "optimizer.projection2candidate();" awaiting the new batcalc api
+ "optimizer.projection2candidate();"
+ "optimizer.deadcode();"
+ "optimizer.projectionpath();"
+ "optimizer.deadcode();"
// "optimizer.oltp();"awaiting the autocommit front-end changes
"optimizer.wlc();"
"optimizer.garbageCollector();",
diff --git a/monetdb5/optimizer/opt_prelude.c b/monetdb5/optimizer/opt_prelude.c
--- a/monetdb5/optimizer/opt_prelude.c
+++ b/monetdb5/optimizer/opt_prelude.c
@@ -145,6 +145,7 @@ str hashRef;
str hgeRef;
str identityRef;
str ifthenelseRef;
+str ilikejoinRef;
str ilikeRef;
str ilikeselectRef;
str ilikethetaselectRef;
@@ -158,6 +159,7 @@ str joinRef;
str jsonRef;
str languageRef;
str leftjoinRef;
+str likejoinRef;
str likeRef;
str likeselectRef;
str likethetaselectRef;
@@ -196,6 +198,7 @@ str oidRef;
str oltpRef;
str openRef;
str optimizerRef;
+str outerjoinRef;
str pack2Ref;
str packIncrementRef;
str packRef;
@@ -442,6 +445,7 @@ void optimizerInit(void)
hgeRef = putName("hge");
identityRef = putName("identity");
ifthenelseRef = putName("ifthenelse");
+ ilikejoinRef = putName("ilikejoin");
ilikeRef = putName("ilike");
ilikeselectRef = putName("ilikeselect");
ilikethetaselectRef = putName("ilikethetaselect");
@@ -455,6 +459,7 @@ void optimizerInit(void)
jsonRef = putName("json");
languageRef= putName("language");
leftjoinRef = putName("leftjoin");
+ likejoinRef = putName("likejoin");
likeRef = putName("like");
likeselectRef = putName("likeselect");
likethetaselectRef = putName("likethetaselect");
@@ -493,6 +498,7 @@ void optimizerInit(void)
oltpRef = putName("oltp");
openRef = putName("open");
optimizerRef = putName("optimizer");
+ outerjoinRef = putName("outerjoin");
pack2Ref = putName("pack2");
packIncrementRef = putName("packIncrement");
packRef = putName("pack");
diff --git a/monetdb5/optimizer/opt_prelude.h b/monetdb5/optimizer/opt_prelude.h
--- a/monetdb5/optimizer/opt_prelude.h
+++ b/monetdb5/optimizer/opt_prelude.h
@@ -146,6 +146,7 @@ mal_export str hashRef;
mal_export str hgeRef;
mal_export str identityRef;
mal_export str ifthenelseRef;
+mal_export str ilikejoinRef;
mal_export str ilikeRef;
mal_export str ilikeselectRef;
mal_export str ilikethetaselectRef;
@@ -160,6 +161,7 @@ mal_export str joinRef;
mal_export str jsonRef;
mal_export str languageRef;
mal_export str leftjoinRef;
+mal_export str likejoinRef;
mal_export str likeRef;
mal_export str likeselectRef;
mal_export str likethetaselectRef;
@@ -199,6 +201,7 @@ mal_export str oidRef;
mal_export str oltpRef;
mal_export str openRef;
mal_export str optimizerRef;
+mal_export str outerjoinRef;
mal_export str pack2Ref;
mal_export str packIncrementRef;
mal_export str packRef;
diff --git a/monetdb5/optimizer/opt_projection2candidate.c
b/monetdb5/optimizer/opt_projection2candidate.c
--- a/monetdb5/optimizer/opt_projection2candidate.c
+++ b/monetdb5/optimizer/opt_projection2candidate.c
@@ -6,18 +6,28 @@
* Copyright 1997 - July 2008 CWI, August 2008 - 2018 MonetDB B.V.
*/
-/* author M.Kersten
- * This optimizer can be used for JIT optimization and moves
- * candidate lists into MAL operations where possible.
- * It should be ran after the candidates optimizer.
- * Specific snippets to be replaced
- * C_1:bat[:oid] := sql.tid(X_0,"sys","t");
- * X_4:bat[:int] := sql.bind(X_0,"sys","t","i",0);
- * X_13 := algebra.projection(C_1,X_4);
- * projection can be avoided
+/* authors M. Kersten ("projection2candidate"), S. Manegold
("projection2candidate")
*
- * A candidate list can be pushed into the calculations
+ * This optimizer is intended to eliminate projections (algebra.projection())
+ * by using the candidate list (first argument of algebra.projection()
+ * directly in the MAL operation that consumes the projection result,
+ * i.e., replace pattern like
+ * X_20 := algebra.projection(C_10, X_10);
+ * (..) := mod.func(.., X_20, ..);
+ * by
+ * (..) := mod.func(.., X_10, .., C_10, ..);
+ * when/where ever the variant of mod.func() with candidate list exists.
+ *
+ * In fact, for actual elimination of no longer used algebra.projection()
+ * calls, this optimizer should be followed by the "deadcode" optimizer!
+ *
+ * This optimizer might merely be a proof-of-concept prototype
+ * that eventually gets replaced by having the SQL-2-MAL code generator
+ * no longer emit algebra.projection() calls, but rather mod.func() calls
+ * that directly use the candidate list (and perform the implicite
+ * projection on-the-fly).
*/
+
#include "monetdb_config.h"
#include "mal_builder.h"
#include "opt_projection2candidate.h"
@@ -33,7 +43,7 @@ OPTprojection2candidateImplementation(Cl
{
int i,actions = 0;
int limit = mb->stop;
- InstrPtr p, q, *old = mb->stmt;
+ InstrPtr p, *old = mb->stmt;
char buf[256];
lng usec = GDKusec();
str msg = MAL_SUCCEED;
@@ -61,23 +71,270 @@ OPTprojection2candidateImplementation(Cl
pushInstruction(mb,old[i]);
break;
}
+#if 0
+// Problems / errors:
+// gdk/gdk_calc.c:5561: BATcalccstsub: Assertion `cand < candend' failed.
+// with TPCH q14
+// inputs not the same size.
+// with TPCH q01, q03, q05, q07, q15, q19
/* case 1
* X_527 := algebra.projection(C_353, X_329);
* X_535 := batcalc.-(100:lng, X_527);
*/
+ /* UNCHNAGED LEGACY CODE (from former "projection2candidate"
optimizer):
+ * - too strict in matching only batcalc.-(const-scalar,COL)
+ * - batcalc.*() implementation with candidate list for
shrinking result
+ * not yet implemented --- well, possibly partly in
batcalc-candidates
+ * branch ...
+ */
if( getModuleId(p) == batcalcRef && *getFunctionId(p) == '-' &&
p->argc == 3 && isVarConstant(mb, getArg(p,1)) ){
- q= getInstrPtr(mb, getVar(mb,getArg(p,2))->updated);
+ InstrPtr q= getInstrPtr(mb,
getVar(mb,getArg(p,2))->updated);
if ( q == 0)
q= getInstrPtr(mb,
getVar(mb,getArg(p,2))->declared);
if( q && getArg(q,0) == getArg(p,2) && getModuleId(q)
== algebraRef && getFunctionId(q) == projectionRef ){
getArg(p,2)= getArg(q,2);
p= pushArgument(mb,p, getArg(q,1));
+ actions++;
OPTDEBUGprojection2candidate(
fprintf(stderr, "#Optimize
projection2candidate case 1\n");
fprintInstruction(stderr,
mb,0,p,LIST_MAL_DEBUG);
)
}
}
+#endif
+#if 0
+ /* case 2:
+ * replace
+ * X_20 := algebra.projection(C_10, X_10);
+ * (..) := group.group[done](X_20);
+ * by
+ * (..) := group.group[done](X_10, C_10);
+ * or
+ * X_20 := algebra.projection(C_10, X_10);
+ * (..) := group.subgroup[done](X_20, X_21 [,X_22, X_23]);
+ * by
+ * (..) := group.subgroup[done](X_10, C_10, X_21 [,X_22,
X_23]));
+ *
+ * the latter fails for now, seemling resulting in an endless
loop ...
+ */
+ if ( getModuleId(p) == groupRef && ( FALSE
+# if 0
+// Problems / errors:
+// gdk/gdk_group.c:521: BATgroup_internal: Assertion `cand < candend' failed.
+// with TPCH q09, q17, q21
+// GDK reported error: BATgroupmin: b and g must be aligned
+// with TPCH q02
+// Can not create object
+// with TPCH q11, q1, q21
+// wrong result
+// with TPCH q04, q05
+ || ( ( getFunctionId(p) == groupRef ||
getFunctionId(p) == groupdoneRef ) && p->argc == p->retc + 1 )
+# endif
+# if 0
+// Problems / errors:
+// gdk/gdk_group.c:521: BATgroup_internal: Assertion `cand < candend' failed.
+// with TPCH q10
+// endless loop
+// with TPCH q01, q03
+// GDK reported error: BATproject: does not match always
+// with TPCH q16
+// Can not create object
+// with TPCH q18
+ || ( ( getFunctionId(p) == subgroupRef || getFunctionId(p)
== subgroupdoneRef ) && ( p->argc == p->retc + 2 || p->argc == p->retc + 4 ) )
+# endif
+ ) )
+ {
+ /* found group.[sub]group[done]() */
+ /* find statement the creates the first argument of
group.[sub]group[done] */
+ InstrPtr q= getInstrPtr(mb,
getVar(mb,getArg(p,p->retc))->updated);
+ if ( q == 0)
+ q= getInstrPtr(mb,
getVar(mb,getArg(p,p->retc))->declared);
+ /* if found, check whether it is algebra.projection() */
+ if( q && q->argc == 3 && getArg(q,0) ==
getArg(p,p->retc) && getModuleId(q) == algebraRef && getFunctionId(q) ==
projectionRef ){
+ if (p->argc == p->retc + 1) {
+ /* group.group[done]():
+ * - simply add candiate list as second
argument
+ */
+ p= pushArgument(mb,p, getArg(q,1)); /*
cand */
+ } else {
+ /* group.subgroup[done]():
+ * - add new argument
+ * - shift all but the first argument
by one
+ * - add candiate list as second
argument
+ */
+ int n = p->argc - 1;
+ p= pushArgument(mb,p, getArg(p,n));
+ for (--n; n > p->retc + 1; n--)
+ getArg(p,n)= getArg(p,n-1);
+ getArg(p,p->retc + 1) = getArg(q,1); /*
cand */
+ }
+ /* replace first argument of
group.[sub]group[done()
+ * by original column, i.e., second argument of
+ * algebra.projection()
+ */
+ getArg(p,p->retc) = getArg(q,2); /* org COL */
+ actions++;
+ OPTDEBUGprojection2candidate(
+ fprintf(stderr, "#Optimize
projection2candidate case 2\n");
+ fprintInstruction(stderr,
mb,0,p,LIST_MAL_DEBUG);
_______________________________________________
checkin-list mailing list
checkin-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/checkin-list
