// getHistoryAttributes
//
// Helper function that traverses the set of root sequence functions
// supplied by the compiler and constructs the set of all of the
// attributes that must be materialized in the history row.
//
void PhysSequence::getHistoryAttributes(const ValueIdSet &sequenceFunctions,
const ValueIdSet &outputFromChild,
ValueIdSet &historyAttributes,
NABoolean addConvNodes,
CollHeap *wHeap,
ValueIdMap *origAttributes) const
{
if(addConvNodes && !origAttributes) {
origAttributes = new (wHeap) ValueIdMap();
}
ValueIdSet children;
for(ValueId valId = sequenceFunctions.init();
sequenceFunctions.next(valId);
sequenceFunctions.advance(valId)) {
if(valId.getItemExpr()->isASequenceFunction()) {
ItemExpr *itmExpr = valId.getItemExpr();
switch(itmExpr->getOperatorType())
{
// The child needs to be in the history row.
//
case ITM_OFFSET:
case ITM_ROWS_SINCE:
case ITM_THIS:
case ITM_NOT_THIS:
// If the child needs to be in the history buffer, then
// add a Convert node to force the value to be moved to the
// history buffer.
if (addConvNodes)
{
itmExpr->child(0) =
addConvNode(itmExpr->child(0), origAttributes, wHeap);
}
historyAttributes += itmExpr->child(0)->getValueId();
break;
// The sequence function needs to be in the history row.
//
case ITM_RUNNING_SUM:
case ITM_RUNNING_COUNT:
case ITM_RUNNING_MIN:
case ITM_RUNNING_MAX:
case ITM_LAST_NOT_NULL:
historyAttributes += itmExpr->getValueId();
break;
/*
// after PhysSequence precode gen OLAP sum and count are already transform,ed into running
// this is used during optimization phase--
case ITM_OLAP_SUM:
case ITM_OLAP_COUNT:
case ITM_OLAP_RANK:
case ITM_OLAP_DRANK:
if (addConvNodes)
{
itmExpr->child(0) =
addConvNode(itmExpr->child(0), origAttributes, wHeap);
}
historyAttributes += itmExpr->child(0)->getValueId();
//historyAttributes += itmExpr->getValueId();
break;
*/
// The child and sequence function need to be in the history row.
//
case ITM_OLAP_MIN:
case ITM_OLAP_MAX:
case ITM_MOVING_MIN:
case ITM_MOVING_MAX:
// If the child needs to be in the history buffer, then
// add a Convert node to force the value to be moved to the
// history buffer.
if (addConvNodes)
{
itmExpr->child(0) =
addConvNode(itmExpr->child(0), origAttributes, wHeap);
}
historyAttributes += itmExpr->child(0)->getValueId();
historyAttributes += itmExpr->getValueId();
break;
case ITM_RUNNING_CHANGE:
if (itmExpr->child(0)->getOperatorType() == ITM_ITEM_LIST)
{
// child is a multi-valued expression
//
ExprValueId treePtr = itmExpr->child(0);
ItemExprTreeAsList changeValues(&treePtr,
ITM_ITEM_LIST,
RIGHT_LINEAR_TREE);
CollIndex nc = changeValues.entries();
ItemExpr *newChild = NULL;
if(addConvNodes) {
newChild = addConvNode(changeValues[nc-1], origAttributes, wHeap);
historyAttributes += newChild->getValueId();
} else {
historyAttributes += changeValues[nc-1]->getValueId();
}
// add each item in the list
//
for (CollIndex i = nc; i > 0; i--)
{
if(addConvNodes) {
ItemExpr *conv
= addConvNode(changeValues[i-1], origAttributes, wHeap);
newChild = new(wHeap) ItemList(conv, newChild);
newChild->synthTypeAndValueId(TRUE);
historyAttributes += conv->getValueId();
} else {
historyAttributes += changeValues[i-1]->getValueId();
}
}
if(addConvNodes) {
itmExpr->child(0) = newChild;
}
}
else
{
// If the child needs to be in the history buffer, then
// add a Convert node to force the value to be moved to the
// history buffer.
if (addConvNodes)
{
itmExpr->child(0) =
addConvNode(itmExpr->child(0), origAttributes, wHeap);
}
historyAttributes += itmExpr->child(0)->getValueId();
}
historyAttributes += itmExpr->getValueId();
break;
default:
CMPASSERT(0);
}
}
// Gather all the children, and if not empty, recurse down to the
// next level of the tree.
//
for(Lng32 i = 0; i < valId.getItemExpr()->getArity(); i++)
{
if (!outputFromChild.contains(valId.getItemExpr()->child(i)->getValueId()))
//!valId.getItemExpr()->child(i)->nodeIsPreCodeGenned())
{
children += valId.getItemExpr()->child(i)->getValueId();
}
}
}
if (NOT children.isEmpty())
{
getHistoryAttributes( children,
outputFromChild,
historyAttributes,
addConvNodes,
wHeap,
origAttributes);
}
} // PhysSequence::getHistoryAttributes
void PhysSequence::computeReadNReturnItems( ValueId topSeqVid,
ValueId vid,
const ValueIdSet &outputFromChild,
CollHeap *wHeap)
{
ItemExpr * itmExpr = vid.getItemExpr();
if (outputFromChild.contains(vid))
{
return;
}
//test if itm_minus and then if negative offset ....
if ( itmExpr->getOperatorType() == ITM_OFFSET &&
((ItmSeqOffset *)itmExpr)->getOffsetConstantValue() < 0)
{
readSeqFunctions() -= topSeqVid;
returnSeqFunctions() += topSeqVid;
readSeqFunctions() += itmExpr->child(0)->castToItemExpr()->getValueId();
return;
}
if (itmExpr->getOperatorType() == ITM_MINUS)
{
ItemExpr * chld0 = itmExpr->child(0)->castToItemExpr();
if ( chld0->getOperatorType() == ITM_OFFSET &&
((ItmSeqOffset *)chld0)->getOffsetConstantValue() <0)
{
readSeqFunctions() -= topSeqVid;
returnSeqFunctions() += topSeqVid;
readSeqFunctions() += chld0->child(0)->castToItemExpr()->getValueId();
ItemExpr * chld1 = itmExpr->child(1)->castToItemExpr();
if (chld1->getOperatorType() == ITM_OFFSET &&
((ItmSeqOffset *)chld1)->getOffsetConstantValue() < 0)
{
readSeqFunctions() += chld1->child(0)->castToItemExpr()->getValueId();
}
else
{
readSeqFunctions() += chld1->getValueId();
}
return;
}
}
if (itmExpr->getOperatorType() == ITM_OLAP_MIN ||
itmExpr->getOperatorType() == ITM_OLAP_MAX)
{
ItmSeqOlapFunction * olap = (ItmSeqOlapFunction *)itmExpr;
if (olap->getframeEnd()>0)
{
readSeqFunctions() -= topSeqVid;
returnSeqFunctions() += topSeqVid;
ItemExpr *newChild = new(wHeap) Convert (itmExpr->child(0)->castToItemExpr());
newChild->synthTypeAndValueId(TRUE);
itmExpr->child(0) = newChild;
readSeqFunctions() += newChild->getValueId();
return;
}
}
if (itmExpr->getOperatorType() == ITM_SCALAR_MIN ||
itmExpr->getOperatorType() == ITM_SCALAR_MAX)
{
ItemExpr * chld0 = itmExpr->child(0)->castToItemExpr();
ItemExpr * chld1 = itmExpr->child(1)->castToItemExpr();
if ((chld0->getOperatorType() == ITM_OLAP_MIN && chld1->getOperatorType() == ITM_OLAP_MIN )||
(chld0->getOperatorType() == ITM_OLAP_MAX && chld1->getOperatorType() == ITM_OLAP_MAX ))
{
ItmSeqOlapFunction * olap0 = (ItmSeqOlapFunction *)chld0;
ItmSeqOlapFunction * olap1 = (ItmSeqOlapFunction *)chld1;
if ( olap1->getframeEnd()>0)
{
CMPASSERT(olap0->getframeEnd()==0);
readSeqFunctions() -= topSeqVid;
returnSeqFunctions() += topSeqVid;
readSeqFunctions() += olap0->getValueId();
ItemExpr *newChild = new(wHeap) Convert (olap1->child(0)->castToItemExpr());
newChild->synthTypeAndValueId(TRUE);
olap1->child(0) = newChild;
readSeqFunctions() += newChild->getValueId();
}
else
{
CMPASSERT(olap1->getframeEnd()==0);
readSeqFunctions() -= topSeqVid;
returnSeqFunctions() += topSeqVid;
readSeqFunctions() += olap1->getValueId();
ItemExpr *newChild = new(wHeap) Convert (olap0->child(0)->castToItemExpr());
newChild->synthTypeAndValueId(TRUE);
olap0->child(0) = newChild;
readSeqFunctions() += newChild->getValueId();
}
return;
}
}
for (Int32 i= 0 ; i < itmExpr->getArity(); i++)
{
ItemExpr * chld= itmExpr->child(i);
computeReadNReturnItems(topSeqVid,
chld->getValueId(),
outputFromChild,
wHeap);
}
}//void PhysSequence::computeReadNReturnItems(ItemExpr * other)
// computeHistoryBuffer
//
// Helper function that traverses the set of root sequence functions
// supplied by the compiler and dynamically determines the size
// of the history buffer.
//
void PhysSequence::computeHistoryRows(const ValueIdSet &sequenceFunctions,//historyIds
Lng32 &computedHistoryRows,
Lng32 &unableToCalculate,
NABoolean &unboundedFollowing,
Lng32 &minFollowingRows,
const ValueIdSet &outputFromChild)
{
ValueIdSet children;
ValueIdSet historyAttributes;
Lng32 value = 0;
for(ValueId valId = sequenceFunctions.init();
sequenceFunctions.next(valId);
sequenceFunctions.advance(valId))
{
if(valId.getItemExpr()->isASequenceFunction())
{
ItemExpr *itmExpr = valId.getItemExpr();
switch(itmExpr->getOperatorType())
{
// THIS and NOT THIS are not dynamically computed
//
case ITM_THIS:
case ITM_NOT_THIS:
break;
// The RUNNING functions and LastNotNull all need to go back just one row.
//
case ITM_RUNNING_SUM:
case ITM_RUNNING_COUNT:
case ITM_RUNNING_MIN:
case ITM_RUNNING_MAX:
case ITM_RUNNING_CHANGE:
case ITM_LAST_NOT_NULL:
computedHistoryRows = MAXOF(computedHistoryRows, 2);
break;
///set to unable to compute for now-- will change later to compte values from frameStart_ and frameEnd_
case ITM_OLAP_SUM:
case ITM_OLAP_COUNT:
case ITM_OLAP_MIN:
case ITM_OLAP_MAX:
case ITM_OLAP_RANK:
case ITM_OLAP_DRANK:
{
if ( !outputFromChild.contains(itmExpr->getValueId()))
{
ItmSeqOlapFunction * olap = (ItmSeqOlapFunction*)itmExpr;
if (olap->isFrameStartUnboundedPreceding()) //(olap->getframeStart() == - INT_MAX)
{
computedHistoryRows = MAXOF(computedHistoryRows, 2);
}
else
{
computedHistoryRows = MAXOF(computedHistoryRows, ABS(olap->getframeStart()) + 2);
}
if (!olap->isFrameEndUnboundedFollowing()) //(olap->getframeEnd() != INT_MAX)
{
computedHistoryRows = MAXOF(computedHistoryRows, ABS(olap->getframeEnd()) + 1);
}
if (olap->isFrameEndUnboundedFollowing()) //(olap->getframeEnd() == INT_MAX)
{
unboundedFollowing = TRUE;
if (olap->getframeStart() > 0)
{
minFollowingRows = ((minFollowingRows > olap->getframeStart()) ?
minFollowingRows : olap->getframeStart());
}
} else if (olap->getframeEnd() > 0)
{
minFollowingRows = ((minFollowingRows > olap->getframeEnd()) ?
minFollowingRows : olap->getframeEnd());
}
}
}
break;
// If 'rows since', we cannot determine how much history is needed.
case ITM_ROWS_SINCE:
unableToCalculate = 1;
break;
// The MOVING and OFFSET functions need to go back as far as the value
// of their second child.
//
// The second argument can be:
// Constant: for these, we can use the constant value to set the upper bound
// for the history buffer.
// ItmScalarMinMax(child0, child1) (with operType = ITM_SCALAR_MIN)
// - if child0 or child1 is a constant, then we can use either one
// to set the upper bound.
case ITM_MOVING_MIN:
case ITM_MOVING_MAX:
case ITM_OFFSET:
for(Lng32 i = 1; i < itmExpr->getArity(); i++)
{
if (itmExpr->child(i)->getOperatorType() != ITM_NOTCOVERED)
{
ItemExpr * exprPtr = itmExpr->child(i);
NABoolean negate;
ConstValue *cv = exprPtr->castToConstValue(negate);
if (cv AND cv->canGetExactNumericValue())
{
Lng32 scale;
Int64 value64 = cv->getExactNumericValue(scale);
if(scale == 0 && value64 >= 0 && value64 < INT_MAX)
{
value64 = (negate ? -value64 : value64);
value = MAXOF((Lng32)value64, value);
}
}
else
{
if (exprPtr->getOperatorType() == ITM_SCALAR_MIN)
{
for(Lng32 j = 0; j < exprPtr->getArity(); j++)
{
if (exprPtr->child(j)->getOperatorType()
!= ITM_NOTCOVERED)
{
ItemExpr * exprPtr1 = exprPtr->child(j);
NABoolean negate1;
ConstValue *cv1 = exprPtr1->castToConstValue(negate1);
if (cv1 AND cv1->canGetExactNumericValue())
{
Lng32 scale1;
Int64 value64_1 = cv1->getExactNumericValue(scale1);
if(scale1 == 0 && value64_1 >= 0 && value64_1 < INT_MAX)
{
value64_1 = (negate1 ? -value64_1 : value64_1);
value = MAXOF((Lng32)value64_1, value);
}
}
}
}
}
} // end of inner else
}// end of if
}// end of for
// Check if the value is greater than zero.
// If it is, then save the value, but first
// increment the returned ConstValue by one.
// Otherwise, the offset or moving value was unable
// to be calculated.
if (value > 0)
{
value++;
computedHistoryRows = MAXOF(computedHistoryRows, value);
value = 0;
}
else
unableToCalculate = 1;
break;
default:
CMPASSERT(0);
}
}
// Gather all the children, and if not empty, recurse down to the
// next level of the tree.
//
for(Lng32 i = 0; i < valId.getItemExpr()->getArity(); i++) {
if (//valId.getItemExpr()->child(i)->getOperatorType() != ITM_NOTCOVERED //old stuff
!outputFromChild.contains(valId.getItemExpr()->child(i)->getValueId()))
{
children += valId.getItemExpr()->child(i)->getValueId();
}
}
}
if (NOT children.isEmpty())
{
computeHistoryRows(children,
computedHistoryRows,
unableToCalculate,
unboundedFollowing,
minFollowingRows,
outputFromChild);
}
} // PhysSequence::computeHistoryRows
// ---------------------------------------------------------------------
// Utility Routine: pickOutputs
//
// From the given ColStatDescList, populate columnStats_ with column
// descriptors that are useful based on the characteristic outputs for
// the group.
//
// Always include in the output the current histograms of the input data,
// and, if the histogram is contained in the required output list, then
// this is a useful histogram and will also be output.
//
// ---------------------------------------------------------------------
void EstLogProp::pickOutputs( ColStatDescList & columnStats,
const EstLogPropSharedPtr& inputEstLogProp,
const ValueIdSet specifiedOutputs,
const ValueIdSet predSet)
{
const ColStatDescList & outerColStatsList = inputEstLogProp->getColStats();
ValueIdSet colsRequiringHistograms = specifiedOutputs;
// (i) see if the selection predicates contain any constant value or a
// constant expression
// (ii) check if there are any columns of this table being joined to some other
// columns, which do not appear as characteristics outputs. There should be
// histograms available for these columns, as these might be needed later.
// This problem was seen for temporary tables created as normal_tables by the
// triggers.
colsRequiringHistograms.addSet(predSet.getColumnsForHistogram());
colStats().setMCSkewedValueLists(columnStats.getMCSkewedValueLists()) ;
NABoolean colStatDescAdded = FALSE;
for (CollIndex i=0; i < columnStats.entries(); i++)
{
// we probably don't need 'em all, but this is the easiest way to
// grab all of the multi-column uec information we'll need later
colStats().insertIntoUecList (columnStats.getUecList()) ;
colStats().setScanRowCountWithoutHint(columnStats.getScanRowCountWithoutHint());
NABoolean found = FALSE;
// Note: The following inserts into a ColStatDescList should not
// have to be deep copies. From this point on, ColStatDescs that
// describe the output of the calling operator are read-only.
ColStatDescSharedPtr colStatDesc = columnStats[i];
// the value-id we're looking for
const ValueId columnId = colStatDesc->getVEGColumn() ;
for (CollIndex j=0 ; j < outerColStatsList.entries() ; j++)
{
if (columnId == outerColStatsList[j]->getVEGColumn() OR
(CmpCommon::context()->showQueryStats()))
{
colStats().insert(colStatDesc) ;
found = TRUE;
if(!colStatDescAdded)
colStatDescAdded = TRUE;
break ; // jump to next ColStatDesc
}
}
// OK, the valueid doesn't match directly -- but there are still a
// couple of things to check in order to verify whether or not we're
// interested in keeping the i'th ColStatDesc ...
ValueId throwaway ; // used by the second clause below
if ( NOT found AND
(columnId != NULL_VALUE_ID) AND
(colsRequiringHistograms.contains (columnId) OR
colsRequiringHistograms.referencesTheGivenValue (columnId, throwaway) OR
columnId.isInvolvedInJoinAndConst() OR
CmpCommon::context()->showQueryStats() )
)
{
colStats().insert(colStatDesc);
found = TRUE;
if(!colStatDescAdded)
colStatDescAdded = TRUE;
}
if (CURRSTMT_OPTDEFAULTS->incorporateSkewInCosting())
{
// if the column is referenced for histogram, but is
// not needed beyond this time , then we shall save its
// max freq, which might be used later in costing if this
// column is a part of the partitioning key
ColStatsSharedPtr stat = colStatDesc->getColStats();
if (!(stat->isVirtualColForHist() ) && NOT found &&
!(stat->isOrigFakeHist() ) )
{
const ValueId col = colStatDesc->getColumn();
ColAnalysis * colAnalysis = col.colAnalysis();
if (colAnalysis)
{
NAColumn * column = stat->getStatColumns()[0];
if (column->isReferencedForHistogram())
{
CostScalar maxFreq = columnStats.getMaxFreq(columnId);
colAnalysis->setMaxFreq(maxFreq);
colAnalysis->setFinalUec(stat->getTotalUec());
colAnalysis->setFinalRC(stat->getRowcount());
}
}
}
}
} // for columnStats.entries()
if(!colStatDescAdded && columnStats.entries() > 0)
colStats().insert(columnStats[0]) ;
} // pickOutputs
