int
OPTjsonImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i, j, limit, slimit;
int bu = 0, br = 0, bj = 0;
str nme;
InstrPtr p,q;
int actions = 0;
InstrPtr *old;
(void) pci;
(void) cntxt;
(void) stk; /* to fool compilers */
old= mb->stmt;
limit= mb->stop;
slimit = mb->ssize;
if ( newMalBlkStmt(mb,mb->stop) < 0)
return 0;
for (i = 0; i < limit; i++) {
p = old[i];
if( getModuleId(p) == sqlRef && getFunctionId(p) == affectedRowsRef) {
q = newStmt(mb, jsonRef, resultSetRef);
q = pushArgument(mb, q, bu);
q = pushArgument(mb, q, br);
q = pushArgument(mb, q, bj);
j = getArg(q,0);
p= getInstrPtr(mb,0);
setVarType(mb,getArg(p,0),TYPE_str);
q = newReturnStmt(mb);
getArg(q,0)= getArg(p,0);
pushArgument(mb,q,j);
continue;
}
if( getModuleId(p) == sqlRef && getFunctionId(p) == rsColumnRef) {
nme = getVarConstant(mb,getArg(p,4)).val.sval;
if (strcmp(nme,"uuid")==0)
bu = getArg(p,7);
if (strcmp(nme,"lng")==0)
br = getArg(p,7);
if (strcmp(nme,"json")==0)
bj = getArg(p,7);
freeInstruction(p);
continue;
}
pushInstruction(mb,p);
}
for(; i<slimit; i++)
if (old[i])
freeInstruction(old[i]);
GDKfree(old);
return actions;
}
int
OPTmatpackImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int v, i, j, limit, slimit;
InstrPtr p,q;
int actions = 0;
InstrPtr *old;
char *packIncrementRef = putName("packIncrement", 13);
(void) pci;
(void) cntxt;
(void) stk; /* to fool compilers */
old= mb->stmt;
limit= mb->stop;
slimit = mb->ssize;
if ( newMalBlkStmt(mb,mb->stop) < 0)
return 0;
for (i = 0; i < limit; i++) {
p = old[i];
if( getModuleId(p) == matRef && getFunctionId(p) == packRef && isaBatType(getArgType(mb,p,1))) {
q = newStmt(mb, matRef, packIncrementRef);
v = getArg(q,0);
setVarType(mb,v,getArgType(mb,p,1));
q = pushArgument(mb, q, getArg(p,1));
q = pushInt(mb,q, p->argc - p->retc);
for ( j = 2; j < p->argc; j++) {
q = newStmt(mb,matRef, packIncrementRef);
q = pushArgument(mb, q, v);
q = pushArgument(mb, q, getArg(p,j));
setVarType(mb,getArg(q,0),getVarType(mb,v));
v = getArg(q,0);
}
getArg(q,0) = getArg(p,0);
freeInstruction(p);
continue;
}
pushInstruction(mb,p);
}
for(; i<slimit; i++)
if (old[i])
freeInstruction(old[i]);
GDKfree(old);
return actions;
}
int
OPTgroupsImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr p)
{
int i, actions=0;
InstrPtr q;
InstrPtr *old, *ref;
int limit,slimit;
(void) cntxt;
(void) stk;
if (varGetProp(mb, getArg(mb->stmt[0], 0), inlineProp) != NULL) {
return 0;
}
/* beware, new variables and instructions are introduced */
ref= (InstrPtr*) GDKzalloc(sizeof(InstrPtr) * mb->vtop); /* to find last assignment */
if ( ref == NULL) {
return 0;
}
old= mb->stmt;
limit= mb->stop;
slimit= mb->ssize;
if ( newMalBlkStmt(mb,mb->ssize) <0) {
GDKfree(ref);
return 0;
}
for (i = 0; i<limit; i++){
p= old[i];
if (getModuleId(p) == groupRef && p->argc == 4 && getFunctionId(p) == subgroupRef ){
setFunctionId(p, multicolumnsRef);
ref[getArg(p,0)] = p;
actions++;
OPTDEBUGgroups {
mnstr_printf(cntxt->fdout,"#new groups instruction\n");
printInstruction(cntxt->fdout,mb, 0, p, LIST_MAL_ALL);
}
}
if (getModuleId(p) == groupRef && p->argc == 5 && getFunctionId(p) == subgroupdoneRef && ref[getArg(p,4)] != NULL){
/*
* Try to expand its argument list with what we have found so far.
* This creates a series of derive paths, many of which will be removed during deadcode elimination.
*/
q= copyInstruction(ref[getArg(p,4)]);
q= pushArgument(mb, q, getArg(p,3));
getArg(q,0) = getArg(p,0);
getArg(q,1) = getArg(p,1);
getArg(q,2) = getArg(p,2);
ref[getArg(q,0)] = q;
freeInstruction(p);
p= q;
OPTDEBUGgroups{
mnstr_printf(cntxt->fdout,"#new groups instruction extension\n");
printInstruction(cntxt->fdout,mb, 0, p, LIST_MAL_ALL);
}
}
static InstrPtr
PushArgument(MalBlkPtr mb, InstrPtr p, int arg, int pos)
{
int i;
p = pushArgument(mb, p, arg); /* push at end */
for (i = p->argc-1; i > pos; i--)
getArg(p, i) = getArg(p, i-1);
getArg(p, pos) = arg;
return p;
}
int
OPTpushselectImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i, j, limit, slimit, actions=0, *vars, push_down_delta = 0, nr_topn = 0, nr_likes = 0;
InstrPtr p, *old;
subselect_t subselects;
memset(&subselects, 0, sizeof(subselects));
if( mb->errors)
return 0;
OPTDEBUGpushselect
mnstr_printf(cntxt->fdout,"#Push select optimizer started\n");
(void) stk;
(void) pci;
vars= (int*) GDKzalloc(sizeof(int)* mb->vtop);
if( vars == NULL)
return 0;
limit = mb->stop;
slimit= mb->ssize;
old = mb->stmt;
/* check for bailout conditions */
for (i = 1; i < limit; i++) {
int lastbat;
p = old[i];
for (j = 0; j<p->retc; j++) {
int res = getArg(p, j);
vars[res] = i;
}
if (getModuleId(p) == algebraRef &&
(getFunctionId(p) == tintersectRef || getFunctionId(p) == tinterRef ||
getFunctionId(p) == tdifferenceRef || getFunctionId(p) == tdiffRef)) {
GDKfree(vars);
return 0;
}
if (getModuleId(p) == algebraRef && getFunctionId(p) == sliceRef)
nr_topn++;
if (isLikeOp(p))
nr_likes++;
if (getModuleId(p) == sqlRef && getFunctionId(p) == deltaRef)
push_down_delta++;
if (getModuleId(p) == sqlRef && getFunctionId(p) == tidRef) { /* rewrite equal table ids */
int sname = getArg(p, 2), tname = getArg(p, 3), s;
for (s = 0; s < subselects.nr; s++) {
InstrPtr q = old[vars[subselects.tid[s]]];
int Qsname = getArg(q, 2), Qtname = getArg(q, 3);
if (no_updates(old, vars, getArg(q,1), getArg(p,1)) &&
((sname == Qsname && tname == Qtname) ||
(0 && strcmp(getVarConstant(mb, sname).val.sval, getVarConstant(mb, Qsname).val.sval) == 0 &&
strcmp(getVarConstant(mb, tname).val.sval, getVarConstant(mb, Qtname).val.sval) == 0))) {
clrFunction(p);
p->retc = 1;
p->argc = 2;
getArg(p, 1) = getArg(q, 0);
break;
}
}
}
lastbat = lastbat_arg(mb, p);
if (isSubSelect(p) && p->retc == 1 &&
/* no cand list */ getArgType(mb, p, lastbat) != newBatType(TYPE_oid, TYPE_oid)) {
int i1 = getArg(p, 1), tid = 0;
InstrPtr q = old[vars[i1]];
/* find the table ids */
while(!tid) {
if (getModuleId(q) == algebraRef && getFunctionId(q) == leftfetchjoinRef) {
int i1 = getArg(q, 1);
InstrPtr s = old[vars[i1]];
if (getModuleId(s) == sqlRef && getFunctionId(s) == tidRef)
tid = getArg(q, 1);
if (s->argc == 2 && s->retc == 1) {
int i1 = getArg(s, 1);
InstrPtr s = old[vars[i1]];
if (getModuleId(s) == sqlRef && getFunctionId(s) == tidRef)
tid = getArg(q, 1);
}
break;
} else if (isMapOp(q) && q->argc >= 2 && isaBatType(getArgType(mb, q, 1))) {
int i1 = getArg(q, 1);
q = old[vars[i1]];
} else if (isMapOp(q) && q->argc >= 3 && isaBatType(getArgType(mb, q, 2))) {
int i2 = getArg(q, 2);
q = old[vars[i2]];
} else {
break;
}
}
if (tid && subselect_add(&subselects, tid, getArg(p, 0)) < 0) {
GDKfree(vars);
return 0;
}
}
/* left hand side */
if ( (GDKdebug & (1<<15)) &&
isMatJoinOp(p) && p->retc == 2) {
int i1 = getArg(p, 2), tid = 0;
InstrPtr q = old[vars[i1]];
/* find the table ids */
while(!tid) {
if (getModuleId(q) == algebraRef && getFunctionId(q) == leftfetchjoinRef) {
int i1 = getArg(q, 1);
InstrPtr s = old[vars[i1]];
if (getModuleId(s) == sqlRef && getFunctionId(s) == tidRef)
tid = getArg(q, 1);
break;
} else if (isMapOp(q) && q->argc >= 2 && isaBatType(getArgType(mb, q, 1))) {
int i1 = getArg(q, 1);
q = old[vars[i1]];
} else if (isMapOp(q) && q->argc >= 3 && isaBatType(getArgType(mb, q, 2))) {
int i2 = getArg(q, 2);
q = old[vars[i2]];
} else {
break;
}
}
if (tid && subselect_add(&subselects, tid, getArg(p, 0)) < 0) {
GDKfree(vars);
return 0;
}
}
/* right hand side */
if ( (GDKdebug & (1<<15)) &&
isMatJoinOp(p) && p->retc == 2) {
int i1 = getArg(p, 3), tid = 0;
InstrPtr q = old[vars[i1]];
/* find the table ids */
while(!tid) {
if (getModuleId(q) == algebraRef && getFunctionId(q) == leftfetchjoinRef) {
int i1 = getArg(q, 1);
InstrPtr s = old[vars[i1]];
if (getModuleId(s) == sqlRef && getFunctionId(s) == tidRef)
tid = getArg(q, 1);
break;
} else if (isMapOp(q) && q->argc >= 2 && isaBatType(getArgType(mb, q, 1))) {
int i1 = getArg(q, 1);
q = old[vars[i1]];
} else if (isMapOp(q) && q->argc >= 3 && isaBatType(getArgType(mb, q, 2))) {
int i2 = getArg(q, 2);
q = old[vars[i2]];
} else {
break;
}
}
if (tid && subselect_add(&subselects, tid, getArg(p, 1)) < 0) {
GDKfree(vars);
return 0;
}
}
}
if ((!subselects.nr && !nr_topn && !nr_likes) || newMalBlkStmt(mb, mb->ssize) <0 ) {
GDKfree(vars);
return 0;
}
pushInstruction(mb,old[0]);
for (i = 1; i < limit; i++) {
p = old[i];
/* rewrite batalgebra.like + subselect -> likesubselect */
if (getModuleId(p) == algebraRef && p->retc == 1 && getFunctionId(p) == subselectRef) {
int var = getArg(p, 1);
InstrPtr q = mb->stmt[vars[var]]; /* BEWARE: the optimizer may not add or remove statements ! */
if (isLikeOp(q)) { /* TODO check if getArg(p, 3) value == TRUE */
InstrPtr r = newInstruction(mb, ASSIGNsymbol);
int has_cand = (getArgType(mb, p, 2) == newBatType(TYPE_oid, TYPE_oid));
int a, anti = (getFunctionId(q)[0] == 'n'), ignore_case = (getFunctionId(q)[anti?4:0] == 'i');
setModuleId(r, algebraRef);
setFunctionId(r, likesubselectRef);
getArg(r,0) = getArg(p,0);
r = pushArgument(mb, r, getArg(q, 1));
if (has_cand)
r = pushArgument(mb, r, getArg(p, 2));
for(a = 2; a<q->argc; a++)
r = pushArgument(mb, r, getArg(q, a));
if (r->argc < (4+has_cand))
r = pushStr(mb, r, ""); /* default esc */
if (r->argc < (5+has_cand))
r = pushBit(mb, r, ignore_case);
if (r->argc < (6+has_cand))
r = pushBit(mb, r, anti);
freeInstruction(p);
p = r;
actions++;
}
}
/* inject table ids into subselect
* s = subselect(c, C1..) => subselect(c, t, C1..)
*/
if (isSubSelect(p) && p->retc == 1) {
int tid = 0;
if ((tid = subselect_find_tids(&subselects, getArg(p, 0))) >= 0) {
int lastbat = lastbat_arg(mb, p);
if (getArgType(mb, p, lastbat) == TYPE_bat) /* empty candidate list bat_nil */
getArg(p, lastbat) = tid;
else
p = PushArgument(mb, p, tid, lastbat+1);
/* make sure to resolve again */
p->token = ASSIGNsymbol;
p->typechk = TYPE_UNKNOWN;
p->fcn = NULL;
p->blk = NULL;
actions++;
}
}
else if ( (GDKdebug & (1<<15)) &&
isMatJoinOp(p) && p->retc == 2
&& !(getFunctionId(p) == joinRef && p->argc > 4)
) {
int ltid = 0, rtid = 0, done = 0;
int range = 0;
if(getFunctionId(p) == joinRef)
range = (p->argc >= 4);
if ((ltid = subselect_find_tids(&subselects, getArg(p, 0))) >= 0 &&
(rtid = subselect_find_tids(&subselects, getArg(p, 1))) >= 0) {
p = PushArgument(mb, p, ltid, 4+range);
p = PushArgument(mb, p, rtid, 5+range);
done = 1;
} else if ((ltid = subselect_find_tids(&subselects, getArg(p, 0))) >= 0) {
p = PushArgument(mb, p, ltid, 4+range);
p = PushNil(mb, p, 5+range, TYPE_bat);
done = 1;
} else if ((rtid = subselect_find_tids(&subselects, getArg(p, 1))) >= 0) {
p = PushNil(mb, p, 4+range, TYPE_bat);
p = PushArgument(mb, p, rtid, 5+range);
done = 1;
}
if (done) {
if(getFunctionId(p) == antijoinRef)
p = pushInt(mb, p, JOIN_NE);
p = pushBit(mb, p, FALSE); /* do not match nils */
p = pushNil(mb, p, TYPE_lng); /* no estimate */
/* TODO join* -> subjoin* */
if(getFunctionId(p) == joinRef)
getFunctionId(p) = subjoinRef;
else if(getFunctionId(p) == antijoinRef)
getFunctionId(p) = subthetajoinRef;
else if(getFunctionId(p) == thetajoinRef)
getFunctionId(p) = subthetajoinRef;
else if(getFunctionId(p) == bandjoinRef)
getFunctionId(p) = subbandjoinRef;
/* make sure to resolve again */
p->token = ASSIGNsymbol;
p->typechk = TYPE_UNKNOWN;
p->fcn = NULL;
p->blk = NULL;
actions++;
}
}
/* Leftfetchjoins involving rewriten table ids need to be flattend
* l = leftfetchjoin(t, c); => l = c;
* and
* l = leftfetchjoin(s, ntids); => l = s;
*/
else if (getModuleId(p) == algebraRef && getFunctionId(p) == leftfetchjoinRef) {
int var = getArg(p, 1);
if (subselect_find_subselect(&subselects, var) > 0) {
InstrPtr q = newAssignment(mb);
getArg(q, 0) = getArg(p, 0);
(void) pushArgument(mb, q, getArg(p, 2));
actions++;
freeInstruction(p);
continue;
} else { /* deletes/updates use table ids */
int var = getArg(p, 2);
InstrPtr q = mb->stmt[vars[var]]; /* BEWARE: the optimizer may not add or remove statements ! */
if (q->token == ASSIGNsymbol) {
var = getArg(q, 1);
q = mb->stmt[vars[var]];
}
if (subselect_find_subselect(&subselects, var) > 0) {
InstrPtr qq = newAssignment(mb);
/* TODO: check result */
getArg(qq, 0) = getArg(p, 0);
(void) pushArgument(mb, qq, getArg(p, 1));
actions++;
freeInstruction(p);
continue;
}
/* c = sql.delta(b,uid,uval,ins);
* l = leftfetchjoin(x, c);
* into
* l = sql.projectdelta(x,b,uid,uval,ins);
*/
else if (getModuleId(q) == sqlRef && getFunctionId(q) == deltaRef && q->argc == 5) {
q = copyInstruction(q);
setFunctionId(q, projectdeltaRef);
getArg(q, 0) = getArg(p, 0);
q = PushArgument(mb, q, getArg(p, 1), 1);
freeInstruction(p);
p = q;
actions++;
}
}
}
pushInstruction(mb,p);
}
for (; i<limit; i++)
if (old[i])
pushInstruction(mb,old[i]);
for (; i<slimit; i++)
if (old[i])
freeInstruction(old[i]);
GDKfree(old);
if (!push_down_delta) {
GDKfree(vars);
return actions;
}
/* now push selects through delta's */
limit = mb->stop;
slimit= mb->ssize;
old = mb->stmt;
if (newMalBlkStmt(mb, mb->stop+(5*push_down_delta)) <0 ) {
mb->stmt = old;
GDKfree(vars);
return actions;
}
pushInstruction(mb,old[0]);
for (i = 1; i < limit; i++) {
int lastbat;
p = old[i];
for (j = 0; j<p->retc; j++) {
int res = getArg(p, j);
vars[res] = i;
}
/* push subslice under projectdelta */
if (isSlice(p) && p->retc == 1) {
int var = getArg(p, 1);
InstrPtr q = old[vars[var]];
if (getModuleId(q) == sqlRef && getFunctionId(q) == projectdeltaRef) {
InstrPtr r = copyInstruction(p);
InstrPtr s = copyInstruction(q);
ValRecord cst;
/* slice the candidates */
setFunctionId(r, sliceRef);
getArg(r, 0) = newTmpVariable(mb, newBatType(TYPE_oid, TYPE_oid));
getArg(r, 1) = getArg(s, 1);
cst.vtype = getArgType(mb, r, 2);
cst.val.wval = 0;
getArg(r, 2) = defConstant(mb, cst.vtype, &cst); /* start from zero */
pushInstruction(mb,r);
/* dummy result for the old q, will be removed by deadcode optimizer */
getArg(q, 0) = newTmpVariable(mb, getArgType(mb, q, 0));
getArg(s, 1) = getArg(r, 0); /* use result of subslice */
pushInstruction(mb, s);
}
}
/* c = delta(b, uid, uvl, ins)
* s = subselect(c, C1..)
*
* nc = subselect(b, C1..)
* ni = subselect(ins, C1..)
* nu = subselect(uvl, C1..)
* s = subdelta(nc, uid, nu, ni);
*
* doesn't handle Xsubselect(x, .. z, C1.. cases) ie multicolumn selects
*/
lastbat = lastbat_arg(mb, p);
if (isSubSelect(p) && p->retc == 1 && lastbat == 2) {
int var = getArg(p, 1);
InstrPtr q = old[vars[var]];
if (q->token == ASSIGNsymbol) {
var = getArg(q, 1);
q = old[vars[var]];
}
if (getModuleId(q) == sqlRef && getFunctionId(q) == deltaRef) {
InstrPtr r = copyInstruction(p);
InstrPtr s = copyInstruction(p);
InstrPtr t = copyInstruction(p);
InstrPtr u = copyInstruction(q);
getArg(r, 0) = newTmpVariable(mb, newBatType(TYPE_oid, TYPE_oid));
getArg(r, 1) = getArg(q, 1); /* column */
pushInstruction(mb,r);
getArg(s, 0) = newTmpVariable(mb, newBatType(TYPE_oid, TYPE_oid));
getArg(s, 1) = getArg(q, 3); /* updates */
s = ReplaceWithNil(mb, s, 2, TYPE_bat); /* no candidate list */
setArgType(mb, s, 2, newBatType(TYPE_oid,TYPE_oid));
/* make sure to resolve again */
s->token = ASSIGNsymbol;
s->typechk = TYPE_UNKNOWN;
s->fcn = NULL;
s->blk = NULL;
pushInstruction(mb,s);
getArg(t, 0) = newTmpVariable(mb, newBatType(TYPE_oid, TYPE_oid));
getArg(t, 1) = getArg(q, 4); /* inserts */
pushInstruction(mb,t);
setFunctionId(u, subdeltaRef);
getArg(u, 0) = getArg(p,0);
getArg(u, 1) = getArg(r,0);
getArg(u, 2) = getArg(p,2); /* pre-cands */
getArg(u, 3) = getArg(q,2); /* update ids */
getArg(u, 4) = getArg(s,0);
u = pushArgument(mb, u, getArg(t,0));
pushInstruction(mb,u);
freeInstruction(p);
continue;
}
}
pushInstruction(mb,p);
}
for (; i<limit; i++)
if (old[i])
pushInstruction(mb,old[i]);
GDKfree(vars);
GDKfree(old);
return actions;
}
static int
OPTsql_appendImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
InstrPtr *old = NULL;
int i, limit, slimit, actions = 0;
(void) pci; /* Tell compilers that we know that we do not */
(void) stk; /* use these function parameters, here. */
/* In general, a MAL optimizer transforms a given MAL program into a
* modified one by sequentially walking through the given program
* and concurrently creating a new one from scratch by
* (1) copying statements as is, modified, or in a different order,
* or (2) omitting statements or (3) introducing new statements.
*/
/* check for logical error: mb must never be NULL */
assert (mb != NULL);
/* save the old stage of the MAL block */
old = mb->stmt;
limit= mb->stop;
slimit = mb->ssize;
/* initialize the statement list. Notice, the symbol table remains intact */
if (newMalBlkStmt(mb, mb->ssize) < 0)
return 0;
/* the plan signature can be copied safely */
pushInstruction(mb, old[0]);
/* iterate over the instructions of the input MAL program */
for (i = 1; i < limit; i++) {
InstrPtr p = old[i];
/* check for
* v3 := sql.append( ..., ..., ..., ..., v0 );
*/
if (getModuleId(p) == sqlRef &&
getFunctionId(p) == appendRef &&
p->argc > 5 &&
p->retc == 1 &&
isaBatType(getArgType(mb, p, 5))) {
/* found
* v3 := sql.append( ..., ..., ..., ..., v0 );
*/
int j = 0, k = 0;
InstrPtr q1 = NULL, q2 = NULL;
bit found = FALSE;
/* check whether next is
* v4 := aggr.count(v0);
*/
if (i+1 < limit) {
InstrPtr q = old[i+1];
if (getModuleId(q) == aggrRef &&
getFunctionId(q) == countRef &&
q->argc == 2 &&
q->retc == 1 &&
getArg(q, 1) == getArg(p, 5)) {
/* found
* v3 := sql.append( ..., ..., ..., ..., v0 );
* v4 := aggr.count(v0);
*/
/* issue/execute
* v4 := aggr.count(v0);
* before
* v3 := sql.append( ..., ..., ..., ..., v0 );
*/
pushInstruction(mb, q);
q1 = q;
i++;
actions++; /* to keep track if anything has been done */
}
}
/* look for
* v5 := ... v0 ...;
*/
/* an expensive loop, better would be to remember that v0
* has a different role. A typical method is to keep a
* map from variable -> instruction where it was
* detected. Then you can check each assignment for use of
* v0
*/
for (j = i+1; !found && j < limit; j++)
for (k = old[j]->retc; !found && k < old[j]->argc; k++)
found = (getArg(old[j], k) == getArg(p, 5));
if (found) {
/* replace
* v3 := sql.append( ..., ..., ..., ..., v0 );
* with
* v1 := aggr.count( v0 );
* v2 := algebra.slice( v0, 0, v1 );
* v3 := sql.append( ..., ..., ..., ..., v2 );
*/
/* push new v1 := aggr.count( v0 ); unless already available */
if (q1 == NULL) {
/* use mal_builder.h primitives
* q1 = newStmt(mb, aggrRef,countRef);
* setArgType(mb,q1,TYPE_wrd) */
/* it will be added to the block and even my
* re-use MAL instructions */
q1 = newInstruction(mb,ASSIGNsymbol);
getArg(q1,0) = newTmpVariable(mb, TYPE_wrd);
setModuleId(q1, aggrRef);
setFunctionId(q1, countRef);
q1 = pushArgument(mb, q1, getArg(p, 5));
pushInstruction(mb, q1);
}
/* push new v2 := algebra.slice( v0, 0, v1 ); */
/* use mal_builder.h primitives
* q1 = newStmt(mb, algebraRef,sliceRef); */
q2 = newInstruction(mb,ASSIGNsymbol);
getArg(q2,0) = newTmpVariable(mb, TYPE_any);
setModuleId(q2, algebraRef);
setFunctionId(q2, sliceRef);
q2 = pushArgument(mb, q2, getArg(p, 5));
q2 = pushWrd(mb, q2, 0);
q2 = pushArgument(mb, q2, getArg(q1, 0));
pushInstruction(mb, q2);
/* push modified v3 := sql.append( ..., ..., ..., ..., v2 ); */
getArg(p, 5) = getArg(q2, 0);
pushInstruction(mb, p);
actions++;
continue;
}
}
pushInstruction(mb, p);
if (p->token == ENDsymbol) break;
}
/* We would like to retain everything from the ENDsymbol
* up to the end of the plan, because after the ENDsymbol
* the remaining optimizer steps are stored.
*/
for(i++; i<limit; i++)
if (old[i])
pushInstruction(mb, old[i]);
/* any remaining MAL instruction records are removed */
for(; i<slimit; i++)
if (old[i])
freeInstruction(old[i]);
GDKfree(old);
/* for statistics we return if/how many patches have been made */
DEBUGoptimizers
mnstr_printf(cntxt->fdout,"#opt_sql_append: %d statements added\n",
actions);
return actions;
}
/*
* The trace operation collects the events in the BATs
* and creates a secondary result set upon termination
* of the query.
*/
static void
SQLsetTrace(backend *be, Client cntxt, bit onoff)
{
InstrPtr q, resultset;
InstrPtr tbls, cols, types, clen, scale;
MalBlkPtr mb = cntxt->curprg->def;
int k;
(void) be;
if (onoff) {
(void) newStmt(mb, "profiler", "start");
initTrace();
} else {
(void) newStmt(mb, "profiler", "stop");
/* cook a new resultSet instruction */
resultset = newInstruction(mb,ASSIGNsymbol);
setModuleId(resultset, sqlRef);
setFunctionId(resultset, resultSetRef);
getArg(resultset,0)= newTmpVariable(mb,TYPE_int);
/* build table defs */
tbls = newStmt(mb,batRef, newRef);
setVarType(mb, getArg(tbls,0), newBatType(TYPE_oid, TYPE_str));
tbls = pushType(mb, tbls, TYPE_oid);
tbls = pushType(mb, tbls, TYPE_str);
resultset= pushArgument(mb,resultset, getArg(tbls,0));
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q,getArg(tbls,0));
q= pushStr(mb,q,".trace");
k= getArg(q,0);
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q,k);
q= pushStr(mb,q,".trace");
/* build colum defs */
cols = newStmt(mb,batRef, newRef);
setVarType(mb, getArg(cols,0), newBatType(TYPE_oid, TYPE_str));
cols = pushType(mb, cols, TYPE_oid);
cols = pushType(mb, cols, TYPE_str);
resultset= pushArgument(mb,resultset, getArg(cols,0));
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q,getArg(cols,0));
q= pushStr(mb,q,"usec");
k= getArg(q,0);
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q, getArg(cols,0));
q= pushStr(mb,q,"statement");
/* build type defs */
types = newStmt(mb,batRef, newRef);
setVarType(mb, getArg(types,0), newBatType(TYPE_oid, TYPE_str));
types = pushType(mb, types, TYPE_oid);
types = pushType(mb, types, TYPE_str);
resultset= pushArgument(mb,resultset, getArg(types,0));
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q, getArg(types,0));
q= pushStr(mb,q,"bigint");
k= getArg(q,0);
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q, k);
q= pushStr(mb,q,"clob");
/* build scale defs */
clen = newStmt(mb,batRef, newRef);
setVarType(mb, getArg(clen,0), newBatType(TYPE_oid, TYPE_int));
clen = pushType(mb, clen, TYPE_oid);
clen = pushType(mb, clen, TYPE_int);
resultset= pushArgument(mb,resultset, getArg(clen,0));
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q, getArg(clen,0));
q= pushInt(mb,q,64);
k= getArg(q,0);
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q, k);
q= pushInt(mb,q,0);
/* build scale defs */
scale = newStmt(mb,batRef, newRef);
setVarType(mb, getArg(scale,0), newBatType(TYPE_oid, TYPE_int));
scale = pushType(mb, scale, TYPE_oid);
scale = pushType(mb, scale, TYPE_int);
resultset= pushArgument(mb,resultset, getArg(scale,0));
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb,q, getArg(scale,0));
q= pushInt(mb,q,0);
k= getArg(q,0);
q= newStmt(mb,batRef,appendRef);
q= pushArgument(mb, q, k);
q= pushInt(mb,q,0);
/* add the ticks column */
q = newStmt(mb, profilerRef, "getTrace");
q = pushStr(mb, q, putName("usec",4));
resultset= pushArgument(mb,resultset, getArg(q,0));
/* add the stmt column */
q = newStmt(mb, profilerRef, "getTrace");
q = pushStr(mb, q, putName("stmt",4));
resultset= pushArgument(mb,resultset, getArg(q,0));
pushInstruction(mb,resultset);
}
}
/*
* The generic solution to the multiplex operators is to translate
* them to a MAL loop.
* The call optimizer.multiplex(MOD,FCN,A1,...An) introduces the following code
* structure:
*
* @verbatim
* A1rev:=bat.reverse(A1);
* resB:= bat.new(A1);
* barrier (h,t):= iterator.new(A1);
* $1:= algebra.fetch(A1,h);
* $2:= A2; # in case of constant?
* ...
* cr:= MOD.FCN($1,...,$n);
* y:=algebra.fetch(A1rev,h);
* bat.insert(resB,y,cr);
* redo (h,t):= iterator.next(A1);
* end h;
* @end verbatim
*
* The algorithm consists of two phases: phase one deals with
* collecting the relevant information, phase two is the actual
* code construction.
*/
static str
OPTexpandMultiplex(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i = 2, resB, iter = 0, cr;
int hvar, tvar;
int x, y;
str mod, fcn;
int *alias;
InstrPtr q;
int ht, tt;
(void) cntxt;
(void) stk;
ht = getHeadType(getArgType(mb, pci, 0));
if (ht != TYPE_oid)
throw(MAL, "optimizer.multiplex", "Target head type is missing");
tt = getTailType(getArgType(mb, pci, 0));
if (tt== TYPE_any)
throw(MAL, "optimizer.multiplex", "Target tail type is missing");
if (isAnyExpression(getArgType(mb, pci, 0)))
throw(MAL, "optimizer.multiplex", "Target type is missing");
mod = VALget(&getVar(mb, getArg(pci, 1))->value);
mod = putName(mod,strlen(mod));
fcn = VALget(&getVar(mb, getArg(pci, 2))->value);
fcn = putName(fcn,strlen(fcn));
/* search the iterator bat */
for (i = 3; i < pci->argc; i++)
if (isaBatType(getArgType(mb, pci, i))) {
iter = getArg(pci, i);
if (getHeadType(getVarType(mb,iter)) != TYPE_oid)
throw(MAL, "optimizer.multiplex", "Iterator BAT is not OID-headed");
break;
}
if( i == pci->argc)
throw(MAL, "optimizer.multiplex", "Iterator BAT type is missing");
OPTDEBUGmultiplex {
mnstr_printf(cntxt->fdout,"#calling the optimize multiplex script routine\n");
printFunction(cntxt->fdout,mb, 0, LIST_MAL_ALL );
mnstr_printf(cntxt->fdout,"#multiplex against operator %d %s\n",iter, getTypeName(getVarType(mb,iter)));
printInstruction(cntxt->fdout,mb, 0, pci,LIST_MAL_ALL);
}
/*
* Beware, the operator constant (arg=1) is passed along as well,
* because in the end we issue a recursive function call that should
* find the actual arguments at the proper place of the callee.
*/
alias= (int*) GDKmalloc(sizeof(int) * pci->maxarg);
if (alias == NULL)
return NULL;
/* x := bat.reverse(A1); */
x = newTmpVariable(mb, newBatType(getTailType(getVarType(mb,iter)),
getHeadType(getVarType(mb,iter))));
q = newFcnCall(mb, batRef, reverseRef);
getArg(q, 0) = x;
q = pushArgument(mb, q, iter);
/* resB := new(refBat) */
q = newFcnCall(mb, batRef, newRef);
resB = getArg(q, 0);
setVarType(mb, getArg(q, 0), newBatType(ht, tt));
q = pushType(mb, q, ht);
q = pushType(mb, q, tt);
/* barrier (h,r) := iterator.new(refBat); */
q = newFcnCall(mb, iteratorRef, newRef);
q->barrier = BARRIERsymbol;
hvar = newTmpVariable(mb, TYPE_any);
getArg(q,0) = hvar;
tvar = newTmpVariable(mb, TYPE_any);
q= pushReturn(mb, q, tvar);
(void) pushArgument(mb,q,iter);
/* $1:= algebra.fetch(Ai,h) or constant */
alias[i] = tvar;
for (i++; i < pci->argc; i++)
if (isaBatType(getArgType(mb, pci, i))) {
q = newFcnCall(mb, algebraRef, "fetch");
alias[i] = newTmpVariable(mb, getTailType(getArgType(mb, pci, i)));
getArg(q, 0) = alias[i];
q= pushArgument(mb, q, getArg(pci, i));
(void) pushArgument(mb, q, hvar);
}
/* cr:= mod.CMD($1,...,$n); */
q = newFcnCall(mb, mod, fcn);
cr = getArg(q, 0) = newTmpVariable(mb, TYPE_any);
for (i = 3; i < pci->argc; i++)
if (isaBatType(getArgType(mb, pci, i))) {
q= pushArgument(mb, q, alias[i]);
} else {
q = pushArgument(mb, q, getArg(pci, i));
}
/* y := algebra.fetch(x,h); */
y = newTmpVariable(mb, getHeadType(getVarType(mb,iter)));
q = newFcnCall(mb, algebraRef, "fetch");
getArg(q, 0) = y;
q = pushArgument(mb, q, x);
q = pushArgument(mb, q, hvar);
/* insert(resB,h,cr);
not append(resB, cr); the head type (oid) may dynamically change */
q = newFcnCall(mb, batRef, insertRef);
q= pushArgument(mb, q, resB);
q= pushArgument(mb, q, y);
(void) pushArgument(mb, q, cr);
/* redo (h,r):= iterator.next(refBat); */
q = newFcnCall(mb, iteratorRef, nextRef);
q->barrier = REDOsymbol;
getArg(q,0) = hvar;
q= pushReturn(mb, q, tvar);
(void) pushArgument(mb,q,iter);
q = newAssignment(mb);
q->barrier = EXITsymbol;
getArg(q,0) = hvar;
(void) pushReturn(mb, q, tvar);
q = newAssignment(mb);
getArg(q, 0) = getArg(pci, 0);
(void) pushArgument(mb, q, resB);
GDKfree(alias);
return MAL_SUCCEED;
}
int
OPTquerylogImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i, limit, slimit;
InstrPtr p = 0, *old= mb->stmt, q,r;
int argc, io, user,nice,sys,idle,iowait,load, arg, start,finish, name;
int xtime=0, rtime = 0, tuples=0;
InstrPtr defineQuery = NULL;
// query log needed?
if ( !QLOGisset() )
return 0;
(void) pci;
(void) stk; /* to fool compilers */
(void) cntxt;
/* gather information */
for (i = 1; i < mb->stop; i++) {
p = getInstrPtr(mb,i);
if ( getModuleId(p) && idcmp(getModuleId(p), "querylog") == 0 && idcmp(getFunctionId(p),"define")==0){
defineQuery= p;
getVarConstant(mb,getArg(p,3)).val.lval = GDKusec()-getVarConstant(mb,getArg(p,3)).val.lval ;
}
}
if ( defineQuery == NULL)
/* nothing to do */
return 0;
limit= mb->stop;
slimit= mb->ssize;
if ( newMalBlkStmt(mb, mb->ssize) < 0)
return 0;
pushInstruction(mb, old[0]);
/* run the querylog.define operation */
defineQuery = copyInstruction(defineQuery);
setFunctionId(defineQuery, insertRef);
getArg(defineQuery,0) = newTmpVariable(mb,TYPE_any);
defineQuery->token = ASSIGNsymbol;
setModuleId(defineQuery,querylogRef);
/* collect the initial statistics */
q = newStmt(mb, "clients", "getUsername");
name= getArg(q,0)= newVariable(mb,GDKstrdup("name"),TYPE_str);
defineQuery = pushArgument(mb,defineQuery,name);
q = newStmt(mb, "mtime", "current_timestamp");
start= getArg(q,0)= newVariable(mb,GDKstrdup("start"),TYPE_timestamp);
defineQuery = pushArgument(mb,defineQuery,start);
pushInstruction(mb, defineQuery);
q = newStmt1(mb, sqlRef, "argRecord");
for ( argc=1; argc < old[0]->argc; argc++)
q = pushArgument(mb, q, getArg(old[0],argc));
arg= getArg(q,0)= newVariable(mb,GDKstrdup("args"),TYPE_str);
q = newStmt(mb, "alarm", "usec");
xtime = getArg(q,0)= newVariable(mb,GDKstrdup("xtime"),TYPE_lng);
user = newVariable(mb,GDKstrdup("user"),TYPE_lng);
nice = newVariable(mb,GDKstrdup("nice"),TYPE_lng);
sys = newVariable(mb,GDKstrdup("sys"),TYPE_lng);
idle = newVariable(mb,GDKstrdup("idle"),TYPE_lng);
iowait = newVariable(mb,GDKstrdup("iowait"),TYPE_lng);
q = newStmt(mb, "profiler", "cpustats");
q->retc= q->argc =0;
q = pushReturn(mb,q,user);
q = pushReturn(mb,q,nice);
q = pushReturn(mb,q,sys);
q = pushReturn(mb,q,idle);
q = pushReturn(mb,q,iowait);
q = newAssignment(mb);
tuples= getArg(q,0) = newVariable(mb,GDKstrdup("tuples"),TYPE_wrd);
(void) pushWrd(mb,q,1);
for (i = 1; i < limit; i++) {
p = old[i];
if (getModuleId(p)==sqlRef &&
(idcmp(getFunctionId(p),"exportValue")==0 ||
idcmp(getFunctionId(p),"exportResult")==0 ) ) {
q = newStmt(mb, "alarm", "usec");
r = newStmt1(mb, calcRef, "-");
r = pushArgument(mb, r, getArg(q,0));
r = pushArgument(mb, r, xtime);
getArg(r,0)=xtime;
q = newStmt(mb, "alarm", "usec");
rtime= getArg(q,0)= newVariable(mb,GDKstrdup("rtime"),TYPE_lng);
pushInstruction(mb,p);
continue;
}
if ( getModuleId(p) == sqlRef && idcmp(getFunctionId(p),"resultSet")==0 && isaBatType(getVarType(mb,getArg(p,3)))){
q = newStmt(mb, "aggr", "count");
getArg(q,0) = tuples;
(void) pushArgument(mb,q, getArg(p,3));
pushInstruction(mb,p);
continue;
}
if ( p->token== ENDsymbol || p->barrier == RETURNsymbol || p->barrier == YIELDsymbol){
if ( rtime == 0){
q = newStmt(mb, "alarm", "usec");
r = newStmt1(mb, calcRef, "-");
r = pushArgument(mb, r, getArg(q,0));
r = pushArgument(mb, r, xtime);
getArg(r,0)=xtime;
q = newStmt(mb, "alarm", "usec");
rtime= getArg(q,0)= newVariable(mb,GDKstrdup("rtime"),TYPE_lng);
}
q = newStmt(mb, "alarm", "usec");
r = newStmt1(mb, calcRef, "-");
r = pushArgument(mb, r, getArg(q,0));
r = pushArgument(mb, r, rtime);
getArg(r,0)=rtime;
/*
* Post execution statistics gathering
*/
q = newStmt(mb, "mtime", "current_timestamp");
finish= getArg(q,0)= newVariable(mb,GDKstrdup("finish"),TYPE_any);
q = newStmt(mb, "profiler", "cpuload");
load = newVariable(mb,GDKstrdup("load"),TYPE_int);
getArg(q,0)= load;
io = newVariable(mb,GDKstrdup("io"),TYPE_int);
q= pushReturn(mb,q,io);
q = pushArgument(mb,q,user);
q = pushArgument(mb,q,nice);
q = pushArgument(mb,q,sys);
q = pushArgument(mb,q,idle);
q = pushArgument(mb,q,iowait);
q = newStmt(mb, querylogRef, "call");
q = pushArgument(mb, q, start);
q = pushArgument(mb, q, finish);
q = pushArgument(mb, q, arg);
q = pushArgument(mb, q, tuples);
q = pushArgument(mb, q, xtime);
q = pushArgument(mb, q, rtime);
q = pushArgument(mb, q, load);
q = pushArgument(mb, q, io);
pushInstruction(mb,p);
continue;
}
pushInstruction(mb,p);
if (p->barrier == YIELDsymbol){
/* the factory yield may return */
q = newStmt(mb, "mtime", "current_timestamp");
start= getArg(q,0)= newVariable(mb,GDKstrdup("start"),TYPE_any);
q = newStmt1(mb, sqlRef, "argRecord");
for ( argc=1; argc < old[0]->argc; argc++)
q = pushArgument(mb, q, getArg(old[0],argc));
arg= getArg(q,0)= newVariable(mb,GDKstrdup("args"),TYPE_str);
q = newAssignment(mb);
q = pushLng(mb,q,0);
q = newAssignment(mb);
q = pushWrd(mb,q,0);
tuples= getArg(q,0)= newVariable(mb,GDKstrdup("tuples"),TYPE_wrd);
newFcnCall(mb,"profiler","setMemoryFlag");
q->argc--;
pushWrd(mb,q,1);
q = newStmt(mb, "alarm", "usec");
xtime = getArg(q,0)= newVariable(mb,GDKstrdup("xtime"),TYPE_lng);
}
}
for( ; i<slimit; i++)
if(old[i])
freeInstruction(old[i]);
GDKfree(old);
return 1;
}
int
OPTpushrangesImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i,j, limit,actions=0;
InstrPtr p, *old;
int x,y,z;
Range range;
if( mb->errors)
return 0;
range= (Range) GDKzalloc(mb->vtop * sizeof(RangeRec));
if (range == NULL)
return 0;
OPTDEBUGpushranges
mnstr_printf(cntxt->fdout,"#Range select optimizer started\n");
(void) stk;
(void) pci;
limit = mb->stop;
old = mb->stmt;
/*
* In phase I we collect information about constants
*/
for (i = 0; i < limit; i++) {
p = old[i];
if( p->barrier)
break; /* end of optimizer */
for(j=p->retc; j< p->argc; j++)
range[getArg(p,j)].used++;
for(j=0; j<p->retc; j++){
range[getArg(p,j)].lastupdate= i;
if( range[getArg(p,j)].lastrange == 0)
range[getArg(p,j)].lastrange= i;
}
if( getModuleId(p)== algebraRef &&
( getFunctionId(p)== selectRef || getFunctionId(p)== uselectRef) ){
/*
* The operation X:= algebra.select(Y,L,H,Li,Hi) is analysed.
* First, we attempt to propagate the range known for Y onto the
* requested range of X. This may lead to smaller range of
* even the conclusion that X is necessarily empty.
* Of course, only under the condition that Y has not been changed by a
* side-effect since it was bound to X.
*/
x= getArg(p,1);
y= getArg(p,2);
if( range[x].lcst && isVarConstant(mb,y) ){
/* merge lowerbound */
if( ATOMcmp( getVarGDKType(mb,y),
VALptr( &getVarConstant(mb,range[x].lcst)),
VALptr( &getVarConstant(mb,y)) ) > 0){
getArg(p,2)= range[x].lcst;
z= range[x].srcvar;
if( getArg(p,1) == x &&
range[z].lastupdate == range[z].lastrange){
getArg(p,1) = z;
actions++;
}
}
y= getArg(p,3);
/* merge higherbound */
if( ATOMcmp( getVarGDKType(mb,y),
VALptr( &getVarConstant(mb,range[x].hcst)),
VALptr( &getVarConstant(mb,y)) ) < 0 ||
ATOMcmp( getVarGDKType(mb,y),
VALptr( &getVarConstant(mb,y)),
ATOMnilptr(getVarType(mb,y)) ) == 0){
getArg(p,3)= range[x].hcst;
z= range[x].srcvar;
if( getArg(p,1) == x && range[z].lastupdate == range[z].lastrange){
getArg(p,1) = z;
actions++;
}
}
}
/*
* The second step is to assign the result of this exercise to the
* result variable.
*/
x= getArg(p,0);
if( isVarConstant(mb, getArg(p,2)) ){
range[x].lcst = getArg(p,2);
range[x].srcvar= getArg(p,1);
range[x].lastupdate= range[x].lastrange = i;
}
if( isVarConstant(mb, getArg(p,3)) ){
range[x].hcst = getArg(p,3);
range[x].srcvar= getArg(p,1);
range[x].lastupdate= range[x].lastrange = i;
}
/*
* If both range bounds are constant, we can also detect empty results.
* It is empty if L> H or when L=H and the bounds are !(true,true).
*/
x= getArg(p,2);
y= getArg(p,3);
if( isVarConstant(mb, x) &&
isVarConstant(mb, y) ){
z =ATOMcmp( getVarGDKType(mb,y),
VALptr( &getVarConstant(mb,x)),
VALptr( &getVarConstant(mb,y)));
x= p->argc > 4;
x= x && isVarConstant(mb,getArg(p,4));
x= x && isVarConstant(mb,getArg(p,5));
x= x && getVarConstant(mb,getArg(p,4)).val.btval;
x= x && getVarConstant(mb,getArg(p,5)).val.btval;
if( z > 0 || (z==0 && p->argc>4 && !x)) {
int var = getArg(p, 0);
wrd zero = 0;
ValRecord v, *vp;
vp = VALset(&v, TYPE_wrd, &zero);
varSetProp(mb, var, rowsProp, op_eq, vp);
/* create an empty replacement */
x = getArgType(mb, p, 1);
p->argc=1;
getModuleId(p)= batRef;
getFunctionId(p)= newRef;
p= pushArgument(mb,p, newTypeVariable(mb, getHeadType(x)));
(void) pushArgument(mb,p, newTypeVariable(mb, getTailType(x)));
actions++;
}
}
}
}
OPTDEBUGpushranges
for(j=0; j< mb->vtop; j++)
if( range[j].used )
printRange(cntxt, mb,range,j);
/*
* Phase II, if we succeeded in pushing constants around and
* changing instructions, we might as well try once more to perform
* aliasRemoval, constantExpression, and pushranges.
*/
GDKfree(range);
return actions;
}
str
OPTvolcanoImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i, limit;
int mvcvar = -1;
int count=0;
InstrPtr p,q, *old = mb->stmt;
char buf[256];
lng usec = GDKusec();
str msg = MAL_SUCCEED;
(void) pci;
(void) cntxt;
(void) stk; /* to fool compilers */
if ( mb->inlineProp )
return MAL_SUCCEED;
limit= mb->stop;
if ( newMalBlkStmt(mb, mb->ssize + 20) < 0)
throw(MAL,"optimizer.volcano", SQLSTATE(HY001) MAL_MALLOC_FAIL);
for (i = 0; i < limit; i++) {
p = old[i];
pushInstruction(mb,p);
if( getModuleId(p) == sqlRef && getFunctionId(p)== mvcRef ){
mvcvar = getArg(p,0);
continue;
}
if( count < MAXdelays && getModuleId(p) == algebraRef ){
if( getFunctionId(p) == selectRef ||
getFunctionId(p) == thetaselectRef ||
getFunctionId(p) == likeselectRef ||
getFunctionId(p) == joinRef
){
q= newInstruction(0,languageRef,blockRef);
setDestVar(q, newTmpVariable(mb,TYPE_any));
q = pushArgument(mb,q,mvcvar);
q = pushArgument(mb,q,getArg(p,0));
mvcvar= getArg(q,0);
pushInstruction(mb,q);
count++;
}
continue;
}
if( count < MAXdelays && getModuleId(p) == groupRef ){
if( getFunctionId(p) == subgroupdoneRef || getFunctionId(p) == groupdoneRef ){
q= newInstruction(0,languageRef,blockRef);
setDestVar(q, newTmpVariable(mb,TYPE_any));
q = pushArgument(mb,q,mvcvar);
q = pushArgument(mb,q,getArg(p,0));
mvcvar= getArg(q,0);
pushInstruction(mb,q);
count++;
}
}
if( getModuleId(p) == sqlRef){
if ( getFunctionId(p) == bindRef ||
getFunctionId(p) == bindidxRef ||
getFunctionId(p)== tidRef ||
getFunctionId(p)== appendRef ||
getFunctionId(p)== updateRef ||
getFunctionId(p)== deleteRef
){
setArg(p,p->retc,mvcvar);
}
}
}
GDKfree(old);
/* Defense line against incorrect plans */
if( count){
chkTypes(cntxt->usermodule, mb, FALSE);
chkFlow(mb);
chkDeclarations(mb);
}
/* keep all actions taken as a post block comment */
usec = GDKusec()- usec;
snprintf(buf,256,"%-20s actions=%2d time=" LLFMT " usec","volcano",count,usec);
newComment(mb,buf);
if( count >= 0)
addtoMalBlkHistory(mb);
return msg;
}
/*
* The generic solution to the multiplex operators is to translate
* them to a MAL loop.
* The call optimizer.multiplex(MOD,FCN,A1,...An) introduces the following code
* structure:
*
* resB:= bat.new(A1);
* barrier (h,t1):= iterator.new(A1);
* t2:= algebra.fetch(A2,h)
* ...
* cr:= MOD.FCN(t1,...,tn);
* bat.append(resB,cr);
* redo (h,t):= iterator.next(A1);
* end h;
*
* The algorithm consists of two phases: phase one deals with
* collecting the relevant information, phase two is the actual
* code construction.
*/
static str
OPTexpandMultiplex(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i = 2, iter = 0;
int hvar, tvar;
str mod, fcn;
int *alias, *resB;
InstrPtr q;
int tt;
int bat = (getModuleId(pci) == batmalRef) ;
//if ( optimizerIsApplied(mb,"multiplex"))
//return 0;
(void) cntxt;
(void) stk;
for (i = 0; i < pci->retc; i++) {
tt = getBatType(getArgType(mb, pci, i));
if (tt== TYPE_any)
throw(MAL, "optimizer.multiplex", SQLSTATE(HY002) "Target tail type is missing");
if (isAnyExpression(getArgType(mb, pci, i)))
throw(MAL, "optimizer.multiplex", SQLSTATE(HY002) "Target type is missing");
}
mod = VALget(&getVar(mb, getArg(pci, pci->retc))->value);
mod = putName(mod);
fcn = VALget(&getVar(mb, getArg(pci, pci->retc+1))->value);
fcn = putName(fcn);
if(mod == NULL || fcn == NULL)
throw(MAL, "optimizer.multiplex", SQLSTATE(HY001) MAL_MALLOC_FAIL);
#ifndef NDEBUG
fprintf(stderr,"#WARNING To speedup %s.%s a bulk operator implementation is needed\n#", mod,fcn);
fprintInstruction(stderr, mb, stk, pci, LIST_MAL_DEBUG);
#endif
/* search the iterator bat */
for (i = pci->retc+2; i < pci->argc; i++)
if (isaBatType(getArgType(mb, pci, i))) {
iter = getArg(pci, i);
break;
}
if( i == pci->argc)
throw(MAL, "optimizer.multiplex", SQLSTATE(HY002) "Iterator BAT type is missing");
#ifdef DEBUG_OPT_MULTIPLEX
{ char *tpenme;
fprintf(stderr,"#calling the optimize multiplex script routine\n");
fprintFunction(stderr,mb, 0, LIST_MAL_ALL );
tpenme = getTypeName(getVarType(mb,iter));
fprintf(stderr,"#multiplex against operator %d %s\n",iter, tpenme);
GDKfree(tpenme);
fprintInstruction(stderr,mb, 0, pci,LIST_MAL_ALL);
}
#endif
/*
* Beware, the operator constant (arg=1) is passed along as well,
* because in the end we issue a recursive function call that should
* find the actual arguments at the proper place of the callee.
*/
alias= (int*) GDKmalloc(sizeof(int) * pci->maxarg);
resB = (int*) GDKmalloc(sizeof(int) * pci->retc);
if (alias == NULL || resB == NULL) {
GDKfree(alias);
GDKfree(resB);
return NULL;
}
/* resB := new(refBat) */
for (i = 0; i < pci->retc; i++) {
q = newFcnCall(mb, batRef, newRef);
resB[i] = getArg(q, 0);
tt = getBatType(getArgType(mb, pci, i));
setVarType(mb, getArg(q, 0), newBatType(tt));
q = pushType(mb, q, tt);
}
/* barrier (h,r) := iterator.new(refBat); */
q = newFcnCall(mb, iteratorRef, newRef);
q->barrier = BARRIERsymbol;
hvar = newTmpVariable(mb, TYPE_any);
getArg(q,0) = hvar;
tvar = newTmpVariable(mb, TYPE_any);
q= pushReturn(mb, q, tvar);
(void) pushArgument(mb,q,iter);
/* $1:= algebra.fetch(Ai,h) or constant */
for (i = pci->retc+2; i < pci->argc; i++) {
if (getArg(pci, i) != iter && isaBatType(getArgType(mb, pci, i))) {
q = newFcnCall(mb, algebraRef, "fetch");
alias[i] = newTmpVariable(mb, getBatType(getArgType(mb, pci, i)));
getArg(q, 0) = alias[i];
q= pushArgument(mb, q, getArg(pci, i));
(void) pushArgument(mb, q, hvar);
}
}
/* cr:= mod.CMD($1,...,$n); */
q = newFcnCall(mb, mod, fcn);
for (i = 0; i < pci->retc; i++) {
int nvar = 0;
if (bat) {
tt = getBatType(getArgType(mb, pci, i));
nvar = newTmpVariable(mb, newBatType(tt));
} else {
nvar = newTmpVariable(mb, TYPE_any);
}
if (i)
q = pushReturn(mb, q, nvar);
else
getArg(q, 0) = nvar;
}
for (i = pci->retc+2; i < pci->argc; i++) {
if (getArg(pci, i) == iter) {
q = pushArgument(mb, q, tvar);
} else if (isaBatType(getArgType(mb, pci, i))) {
q = pushArgument(mb, q, alias[i]);
} else {
q = pushArgument(mb, q, getArg(pci, i));
}
}
for (i = 0; i < pci->retc; i++) {
InstrPtr a = newFcnCall(mb, batRef, appendRef);
a = pushArgument(mb, a, resB[i]);
(void) pushArgument(mb, a, getArg(q,i));
}
/* redo (h,r):= iterator.next(refBat); */
q = newFcnCall(mb, iteratorRef, nextRef);
q->barrier = REDOsymbol;
getArg(q,0) = hvar;
q= pushReturn(mb, q, tvar);
(void) pushArgument(mb,q,iter);
q = newAssignment(mb);
q->barrier = EXITsymbol;
getArg(q,0) = hvar;
(void) pushReturn(mb, q, tvar);
for (i = 0; i < pci->retc; i++) {
q = newAssignment(mb);
getArg(q, 0) = getArg(pci, i);
(void) pushArgument(mb, q, resB[i]);
}
GDKfree(alias);
GDKfree(resB);
return MAL_SUCCEED;
}
int
OPTrecyclerImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr p)
{
int i, j, cnt, tp, c, actions = 0, marks = 0, delta = 0;
Lifespan span;
InstrPtr *old, q;
int limit, updstmt = 0;
char *recycled;
short app_sc = -1, in = 0;
ValRecord cst;
(void) cntxt;
(void) stk;
limit = mb->stop;
old = mb->stmt;
for (i = 1; i < limit; i++) {
p = old[i];
if (getModuleId(p) == sqlRef &&
(getFunctionId(p) == affectedRowsRef ||
getFunctionId(p) == exportOperationRef ||
getFunctionId(p) == appendRef ||
getFunctionId(p) == updateRef ||
getFunctionId(p) == deleteRef))
updstmt = 1;
}
span = setLifespan(mb);
if (span == NULL)
return 0;
/* watch out, newly created instructions may introduce new variables */
recycled = GDKzalloc(sizeof(char) * mb->vtop * 2);
if (recycled == NULL)
return 0;
if (newMalBlkStmt(mb, mb->ssize) < 0) {
GDKfree(recycled);
return 0;
}
pushInstruction(mb, old[0]);
mb->recid = recycleSeq++;
/* create a handle for recycler */
(void) newFcnCall(mb, "recycle", "prelude");
in = 1;
for (i = 1; i < limit; i++) {
p = old[i];
if (hasSideEffects(p, TRUE) || isUpdateInstruction(p) || isUnsafeFunction(p)) {
if (getModuleId(p) == recycleRef) { /*don't inline recycle instr. */
freeInstruction(p);
continue;
}
pushInstruction(mb, p);
/* update instructions are not recycled but monitored*/
if (isUpdateInstruction(p)) {
if (getModuleId(p) == batRef &&
(getArgType(mb, p, 1) == TYPE_bat
|| isaBatType(getArgType(mb, p, 1)))) {
recycled[getArg(p, 1)] = 0;
q = newFcnCall(mb, "recycle", "reset");
pushArgument(mb, q, getArg(p, 1));
actions++;
}
if (getModuleId(p) == sqlRef) {
if (getFunctionId(p) == appendRef) {
if (app_sc >= 0)
continue;
else
app_sc = getArg(p, 2);
}
VALset(&cst, TYPE_int, &delta);
c = defConstant(mb, TYPE_int, &cst);
q = newFcnCall(mb, "recycle", "reset");
pushArgument(mb, q, c);
pushArgument(mb, q, getArg(p, 2));
pushArgument(mb, q, getArg(p, 3));
if (getFunctionId(p) == updateRef)
pushArgument(mb, q, getArg(p, 4));
actions++;
}
}
/* take care of SQL catalog update instructions */
if (getModuleId(p) == sqlRef && getFunctionId(p) == catalogRef) {
tp = *(int *) getVarValue(mb, getArg(p, 1));
if (tp == 22 || tp == 25) {
delta = 2;
VALset(&cst, TYPE_int, &delta);
c = defConstant(mb, TYPE_int, &cst);
q = newFcnCall(mb, "recycle", "reset");
pushArgument(mb, q, c);
pushArgument(mb, q, getArg(p, 2));
if (tp == 25)
pushArgument(mb, q, getArg(p, 3));
actions++;
}
}
continue;
}
if (p->token == ENDsymbol || p->barrier == RETURNsymbol) {
if (in) {
/*
if (updstmt && app_sc >= 0) {
q = newFcnCall(mb, "recycle", "reset");
pushArgument(mb, q, app_sc);
pushArgument(mb, q, app_tbl);
}
*/
(void) newFcnCall(mb, "recycle", "epilogue");
in = 0;
}
pushInstruction(mb, p);
continue;
}
if (p->barrier && p->token != CMDcall) {
/* never save a barrier unless it is a command and side-effect free */
pushInstruction(mb, p);
continue;
}
/* don't change instructions in update statements */
if (updstmt) {
pushInstruction(mb, p);
continue;
}
/* skip simple assignments */
if (p->token == ASSIGNsymbol) {
pushInstruction(mb, p);
continue;
}
if (getModuleId(p) == octopusRef &&
(getFunctionId(p) == bindRef || getFunctionId(p) == bindidxRef)) {
recycled[getArg(p, 0)] = 1;
p->recycle = recycleMaxInterest;
marks++;
}
/* During base table recycling skip marking instructions other than octopus.bind */
if (baseTableMode) {
pushInstruction(mb, p);
continue;
}
/* general rule: all arguments are constants or recycled,
ignore C pointer arguments from mvc */
cnt = 0;
for (j = p->retc; j < p->argc; j++)
if (recycled[getArg(p, j)] || isVarConstant(mb, getArg(p, j))
|| ignoreVar(mb, getArg(p, j)))
cnt++;
if (cnt == p->argc - p->retc) {
OPTDEBUGrecycle {
mnstr_printf(cntxt->fdout, "#recycle instruction\n");
printInstruction(cntxt->fdout, mb, 0, p, LIST_MAL_ALL);
}
marks++;
p->recycle = recycleMaxInterest; /* this instruction is to be monitored */
for (j = 0; j < p->retc; j++)
if (getLastUpdate(span, getArg(p, j)) == i)
recycled[getArg(p, j)] = 1;
}
/*
* The expected gain is largest if we can re-use selections
* on the base tables in SQL. These, however, are marked as
* uselect() calls, which only produce the oid head.
* For cheap types we preselect using select() and re-map uselect() back
* over this temporary.
* For the time being for all possible selects encountered
* are marked for re-use.
*/
/* take care of semantic driven recyling */
/* for selections check the bat argument only
the range is often template parameter*/
if ((getFunctionId(p) == selectRef ||
getFunctionId(p) == antiuselectRef ||
getFunctionId(p) == likeselectRef ||
getFunctionId(p) == likeRef ||
getFunctionId(p) == thetaselectRef) &&
recycled[getArg(p, 1)])
{
p->recycle = recycleMaxInterest;
marks++;
if (getLastUpdate(span, getArg(p, 0)) == i)
recycled[getArg(p, 0)] = 1;
}
if ((getFunctionId(p) == uselectRef || getFunctionId(p) == thetauselectRef)
&& recycled[getArg(p, 1)])
{
if (!ATOMvarsized(getGDKType(getArgType(mb, p, 2)))) {
q = copyInstruction(p);
getArg(q, 0) = newTmpVariable(mb, TYPE_any);
if (getFunctionId(p) == uselectRef)
setFunctionId(q, selectRef);
else
setFunctionId(q, thetaselectRef);
q->recycle = recycleMaxInterest;
marks++;
recycled[getArg(q, 0)] = 1;
pushInstruction(mb, q);
getArg(p, 1) = getArg(q, 0);
setFunctionId(p, projectRef);
p->argc = 2;
}
p->recycle = recycleMaxInterest;
marks++;
if (getLastUpdate(span, getArg(p, 0)) == i)
recycled[getArg(p, 0)] = 1;
}
if (getModuleId(p) == pcreRef) {
if ((getFunctionId(p) == selectRef && recycled[getArg(p, 2)]) ||
(getFunctionId(p) == uselectRef && recycled[getArg(p, 2)])) {
p->recycle = recycleMaxInterest;
marks++;
if (getLastUpdate(span, getArg(p, 0)) == i)
recycled[getArg(p, 0)] = 1;
} else if (getFunctionId(p) == likeuselectRef && recycled[getArg(p, 1)]) {
q = copyInstruction(p);
getArg(q, 0) = newTmpVariable(mb, TYPE_any);
setFunctionId(q, likeselectRef);
q->recycle = recycleMaxInterest;
recycled[getArg(q, 0)] = 1;
pushInstruction(mb, q);
getArg(p, 1) = getArg(q, 0);
setFunctionId(p, projectRef);
setModuleId(p, algebraRef);
p->argc = 2;
p->recycle = recycleMaxInterest;
marks += 2;
if (getLastUpdate(span, getArg(p, 0)) == i)
recycled[getArg(p, 0)] = 1;
}
}
/*
* The sql.bind instructions should be handled carefully
* The delete and update BATs should not be recycled,
* because they may lead to view dependencies that later interferes
* with the transaction commits.
*/
/* enable recycling of delta-bats
if (getModuleId(p) == sqlRef &&
(((getFunctionId(p) == bindRef || getFunctionId(p) == putName("bind_idxbat", 11)) &&
getVarConstant(mb, getArg(p, 5)).val.ival != 0) ||
getFunctionId(p) == binddbatRef)) {
recycled[getArg(p, 0)] = 0;
p->recycle = REC_NO_INTEREST;
}
*/
/*
* The sql.bind instructions should be handled carefully
* The delete and update BATs should not be recycled,
* because they may lead to view dependencies that later interferes
* with the transaction commits.
*/
/* enable recycling of delta-bats
if (getModuleId(p)== sqlRef &&
(((getFunctionId(p)==bindRef || getFunctionId(p) == putName("bind_idxbat",11)) &&
getVarConstant(mb, getArg(p,5)).val.ival != 0) ||
getFunctionId(p)== binddbatRef) ) {
recycled[getArg(p,0)]=0;
p->recycle = REC_NO_INTEREST;
}
*/
pushInstruction(mb, p);
}
str
OPTjsonImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i, j, limit, slimit;
int bu = 0, br = 0, bj = 0;
str nme;
InstrPtr p,q;
int actions = 0;
InstrPtr *old;
char buf[256];
lng usec = GDKusec();
str msg = MAL_SUCCEED;
(void) pci;
(void) cntxt;
(void) stk; /* to fool compilers */
old= mb->stmt;
limit= mb->stop;
slimit = mb->ssize;
if ( newMalBlkStmt(mb,mb->stop) < 0)
throw(MAL,"optimizer.json", SQLSTATE(HY001) MAL_MALLOC_FAIL);
for (i = 0; i < limit; i++) {
p = old[i];
if( getModuleId(p) == sqlRef && getFunctionId(p) == affectedRowsRef) {
q = newInstruction(0, jsonRef, resultSetRef);
q = pushArgument(mb, q, bu);
q = pushArgument(mb, q, br);
q = pushArgument(mb, q, bj);
j = getArg(q,0);
p= getInstrPtr(mb,0);
setDestVar(q, newTmpVariable(mb, TYPE_str));
pushInstruction(mb,p);
q = newInstruction(0, NULL, NULL);
q->barrier = RETURNsymbol;
getArg(q,0)= getArg(p,0);
pushArgument(mb,q,j);
pushInstruction(mb,q);
actions++;
continue;
}
if( getModuleId(p) == sqlRef && getFunctionId(p) == rsColumnRef) {
nme = getVarConstant(mb,getArg(p,4)).val.sval;
if (strcmp(nme,"uuid")==0)
bu = getArg(p,7);
if (strcmp(nme,"lng")==0)
br = getArg(p,7);
if (strcmp(nme,"json")==0)
bj = getArg(p,7);
freeInstruction(p);
actions++;
continue;
}
pushInstruction(mb,p);
}
for(; i<slimit; i++)
if (old[i])
freeInstruction(old[i]);
GDKfree(old);
/* Defense line against incorrect plans */
if( actions > 0){
chkTypes(cntxt->usermodule, mb, FALSE);
chkFlow(mb);
chkDeclarations(mb);
}
/* keep all actions taken as a post block comment */
usec = GDKusec()- usec;
snprintf(buf,256,"%-20s actions=%2d time=" LLFMT " usec","json",actions, usec);
newComment(mb,buf);
if( actions >= 0)
addtoMalBlkHistory(mb);
return msg;
}
int
OPTrecyclerImplementation(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
{
int i, j, cnt, cand, actions = 1, marks = 0;
InstrPtr *old, q,p;
int limit;
char *recycled;
(void) cntxt;
(void) stk;
(void) pci;
limit = mb->stop;
old = mb->stmt;
/* watch out, newly created instructions may introduce new variables */
recycled = GDKzalloc(sizeof(char) * mb->vtop * 2);
if (recycled == NULL)
return 0;
if (newMalBlkStmt(mb, mb->ssize) < 0) {
GDKfree(recycled);
return 0;
}
pushInstruction(mb, old[0]);
for (i = 1; i < limit; i++) {
p = old[i];
if (p->token == ENDsymbol )
break;
/* the first non-dataflow barrier breaks the recycler code*/
if (blockStart(p) && !(getFunctionId(p) && getFunctionId(p) == dataflowRef) )
break;
if ( isUpdateInstruction(p) || hasSideEffects(p,TRUE)){
/* update instructions are not recycled but monitored*/
pushInstruction(mb, p);
if (isUpdateInstruction(p)) {
if (getModuleId(p) == batRef && isaBatType(getArgType(mb, p, 1))) {
q = newFcnCall(mb, "recycle", "reset");
pushArgument(mb, q, getArg(p, 1));
actions++;
}
if (getModuleId(p) == sqlRef) {
q= copyInstruction(p);
getModuleId(q) = recycleRef;
actions++;
}
}
continue;
}
// Not all instruction may be recycled. In particular, we should avoid
// MAL function with implicit/recursive side effects.
// This can not always be detected easily. Likewise, we ignore cheap operations
// Therefore, we use a safe subset to start with
if ( ! (getModuleId(p) == sqlRef || getModuleId(p)== batRef ||
getModuleId(p) == algebraRef || getModuleId(p)==batcalcRef ||
getModuleId(p)== aggrRef || getModuleId(p)== groupRef ||
getModuleId(p)== batstrRef || getModuleId(p)== batmmathRef ||
getModuleId(p)== arrayRef || getModuleId(p)== batmtimeRef ||
getModuleId(p)== batcalcRef || getModuleId(p)== pcreRef ||
getModuleId(p)== mtimeRef || getModuleId(p) == calcRef ||
getModuleId(p)== dateRef || getModuleId(p) == timestampRef ||
getModuleId(p)== matRef )
){
pushInstruction(mb,p);
continue;
}
/* general rule: all arguments should be constants or recycled*/
cnt = 0;
for (j = p->retc; j < p->argc; j++)
if (recycled[getArg(p, j)] || isVarConstant(mb, getArg(p, j)) || isFunctionArgument(mb,getArg(p,j)) )
cnt++;
cand = 0;
for (j =0; j< p->retc; j++)
if (recycled[getArg(p, j)] ==0)
cand++;
if (cnt == p->argc - p->retc && cand == p->retc) {
marks++;
p->recycle = RECYCLING; /* this instruction is to be monitored */
for (j = 0; j < p->retc; j++)
recycled[getArg(p, j)] = 1;
}
pushInstruction(mb, p);
}
for (; i < limit; i++)
pushInstruction(mb, old[i]);
GDKfree(old);
GDKfree(recycled);
mb->recycle = marks > 0;
OPTDEBUGrecycle {
mnstr_printf(cntxt->fdout, "#recycle optimizer: ");
printFunction(cntxt->fdout,mb, 0, LIST_MAL_ALL);
}
return actions + marks;
}
