void
RelSequence::addCancelExpr(CollHeap *wHeap)
{
ItemExpr *cPred = NULL;
if (this->partition().entries() > 0)
{
return;
}
if(cancelExpr().entries() > 0)
{
return;
}
for(ValueId valId = selectionPred().init();
selectionPred().next(valId);
selectionPred().advance(valId))
{
ItemExpr *pred = valId.getItemExpr();
// Look for preds that select a prefix of the sequence.
// Rank() < const; Rank <= const; const > Rank; const >= Rank
ItemExpr *op1 = NULL;
ItemExpr *op2 = NULL;
if(pred->getOperatorType() == ITM_LESS ||
pred->getOperatorType() == ITM_LESS_EQ)
{
op1 = pred->child(0);
op2 = pred->child(1);
}
else if (pred->getOperatorType() == ITM_GREATER ||
pred->getOperatorType() == ITM_GREATER_EQ)
{
op1 = pred->child(1);
op2 = pred->child(0);
}
NABoolean negate;
if (op1 && op2 &&
(op2->getOperatorType() == ITM_CONSTANT ||
op2->getOperatorType() == ITM_DYN_PARAM) &&
(op1->getOperatorType() == ITM_OLAP_RANK ||
op1->getOperatorType() == ITM_OLAP_DRANK ||
(op1->getOperatorType() == ITM_OLAP_COUNT &&
op1->child(0)->getOperatorType() == ITM_CONSTANT &&
!op1->child(0)->castToConstValue(negate)->isNull())))
{
cPred = new(wHeap) UnLogic(ITM_NOT, pred);
//break at first occurence
break;
}
}
if(cPred)
{
cPred->synthTypeAndValueId(TRUE);
cancelExpr().insert(cPred->getValueId());
}
}
// index access (both reference and value)
ItemExpr * ItemExprTreeAsList::operator [] (CollIndex i)
{
// think of three different cases:
//
// a) i is out of range (< 0 or >= #entries)
//
// b) the node we are looking for is neither the only nor the last
// node in the backbone
//
// c) we are looking for the last element in the backbone (which
// may be the only element)
//
ItemExpr *aNodePtr = *treePtr_;
Int32 j = (Int32) i; // j may become negative, i may be unsigned
if (j < 0)
return NULL; // case a
if (aNodePtr->getOperatorType() != operator_ AND
j == 0)
return aNodePtr; // case b
while (aNodePtr != NULL AND j >= 0)
{
if (aNodePtr->getOperatorType() == operator_ AND
aNodePtr->getArity() >= 2)
{
if (shape_ == LEFT_LINEAR_TREE)
{
if (j == 0)
aNodePtr = aNodePtr->child(1); // case b
else
aNodePtr = aNodePtr->child(0);
}
else if (shape_ == RIGHT_LINEAR_TREE)
{
if (j == 0)
aNodePtr = aNodePtr->child(0); // case b
else
aNodePtr = aNodePtr->child(1);
}
else
ABORT("can't do bushy trees");
}
j--;
}
// if we are looking for the only element, the while loop
// is not executed at all
return aNodePtr;
}
// is any literal in this expr safely coercible to its target type?
NABoolean ItemExpr::isSafelyCoercible(CacheWA &cwa) const
{
if (cwa.getPhase() >= CmpMain::BIND) {
Int32 arity = getArity();
for (Int32 x = 0; x < arity; x++) {
if (!child(x)->isSafelyCoercible(cwa)) {
return FALSE;
}
}
if (arity == 2) {
// we have to disallow caching of the following types of exprs:
// expr + 123456789012345678901234567890
// expr || 'overlylongstringthatwouldoverflow'
ItemExpr *left = child(0), *right = child(1);
if (left->getOperatorType() == ITM_CONSTANT) {
if (right->getOperatorType() == ITM_CONSTANT) {
// "10 + 1" should be safely coercible
return TRUE;
} else {
return ((ConstValue*)left)->canBeSafelyCoercedTo
(right->getValueId().getType());
}
}
else if (right->getOperatorType() == ITM_CONSTANT) {
return ((ConstValue*)right)->canBeSafelyCoercedTo
(left->getValueId().getType());
}
// else both are nonliterals; fall thru
}
// else nondyadic expr; fall thru
return TRUE;
}
return FALSE;
}
ValueId NormWA::getEquivalentItmSequenceFunction(ValueId newSeqId)
{
ValueId equivId = newSeqId;
ItemExpr *newItem = newSeqId.getItemExpr();
ItmSequenceFunction *newSeq = NULL;
if(newItem->isASequenceFunction()) {
newSeq = (ItmSequenceFunction *)newItem;
}
if(newSeq) {
for(ValueId seqId = allSeqFunctions_.init(); allSeqFunctions_.next(seqId);
allSeqFunctions_.advance(seqId) ){
ItemExpr *seq = seqId.getItemExpr();
if(newSeq->isEquivalentForBinding(seq)){
equivId = seqId;
if(newSeq->origOpType() != seq->origOpType()) {
seq->setOrigOpType(seq->getOperatorType());
}
break;
}
}
}
allSeqFunctions_ += equivId;
//
return equivId;
}
// change literals of a cacheable query into input parameters
ItemExpr* ItemList::normalizeListForCache
(CacheWA& cwa, BindWA& bindWA, ItemList *other)
{
Int32 arity = getArity();
if (cwa.getPhase() >= CmpMain::BIND &&
other && arity == other->getArity()) {
for (Int32 x = 0; x < arity; x++) {
ItemExpr *leftC = child(x), *rightC = other->child(x);
OperatorTypeEnum leftO = leftC->getOperatorType();
OperatorTypeEnum rightO = rightC->getOperatorType();
if (leftO == ITM_BASECOLUMN && rightO == ITM_CONSTANT) {
parameterizeMe(cwa, bindWA, other->child(x), (BaseColumn*)leftC,
(ConstValue*)rightC);
}
else if (rightO == ITM_BASECOLUMN && leftO == ITM_CONSTANT) {
parameterizeMe(cwa, bindWA, child(x), (BaseColumn*)rightC,
(ConstValue*)leftC);
}
else if (leftO == ITM_ITEM_LIST && rightO == ITM_ITEM_LIST) {
child(x) = ((ItemList*)leftC)->normalizeListForCache
(cwa, bindWA, (ItemList*)rightC);
}
}
}
return this;
}
// -----------------------------------------------------------------------
// Given a column list providing identifiers for columns of this table,
// this method returns a list of VEG expressions and/or base columns that
// show the equivalence of base columns with index columns.
// -----------------------------------------------------------------------
void TableDesc::getEquivVEGCols (const ValueIdList& columnList,
ValueIdList &VEGColumnList) const
{
for (CollIndex i=0; i < columnList.entries(); i++)
{
ItemExpr *ie = columnList[i].getItemExpr();
BaseColumn *bc = NULL;
switch (ie->getOperatorType())
{
case ITM_BASECOLUMN:
bc = (BaseColumn *) ie;
break;
case ITM_INDEXCOLUMN:
bc = (BaseColumn *) ((IndexColumn *) ie)->getDefinition().
getItemExpr();
CMPASSERT(bc->getOperatorType() == ITM_BASECOLUMN);
break;
default:
ABORT("Invalid argument to TableDesc::getEquivVEGCols()\n");
}
CMPASSERT(bc->getTableDesc() == this);
VEGColumnList.insert(getColumnVEGList()[bc->getColNumber()]);
}
}
// A transformation method for protecting sequence functions from not
// being evaluated due to short-circuit evaluation.
//
void Case::protectiveSequenceFunctionTransformation(Generator *generator)
{
// Recurse on the children
//
ItemExpr::protectiveSequenceFunctionTransformation(generator);
// Remove the original value id from the node being transformed and
// assign it a new value id.
//
ValueId id = getValueId();
setValueId(NULL_VALUE_ID);
synthTypeAndValueId(TRUE);
// Construct the new subtree.
//
// Case -- force evaluation of all the WHEN, THEN and ELSE parts
//
// CASE(IFE1(W1,T1,IFE2(W2,T2,IFE3(...)))) ==>
// BLOCK(BLOCK(BLOCK(W1,T1),BLOCK(W2,T2)), CASE(...))
//
// Decend the ITM_IF_THEN_ELSE tree pulling out each WHEN and THEN pair.
// Mate each pair with a block and attach them to the protected block,
// which contains all of the WHEN/THEN pairs for the entire tree.
// Also, pull out any CASE operands and attach them to the protected
// block as well.
//
ItemExpr *block = NULL;
ItemExpr *ife = child(0);
for(; (ife != NULL) && (ife->getOperatorType() == ITM_IF_THEN_ELSE);
ife = ife->child(2))
{
ItemExpr *sub = new(generator->wHeap())
ItmBlockFunction(ife->child(0), ife->child(1));
if(block)
block = new(generator->wHeap()) ItmBlockFunction(sub, block);
else
block = sub;
}
// Add the ELSE condition, if any to the protected block
//
if(ife)
block = new(generator->wHeap()) ItmBlockFunction(ife, block);
// Construct the top-level block function. The left child is the protected
// block, which contains all of the expresssions that need to be
// pre-evaluated. This right child is the original case statement.
//
block = new(generator->wHeap()) ItmBlockFunction(block, this);
// Replace the old expression with the new expression for the
// original id
//
id.replaceItemExpr(block);
// Run the new expression through type and value id synthesis
//
block->synthTypeAndValueId(TRUE);
}
ValueIdSet TableDesc::getComputedColumns(NAColumnBooleanFuncPtrT fptr)
{
ValueIdSet computedColumns;
for (CollIndex j=0; j<getClusteringIndex()->getIndexKey().entries(); j++)
{
ItemExpr *ck = getClusteringIndex()->getIndexKey()[j].getItemExpr();
if (ck->getOperatorType() == ITM_INDEXCOLUMN)
ck = ((IndexColumn *) ck)->getDefinition().getItemExpr();
CMPASSERT(ck->getOperatorType() == ITM_BASECOLUMN);
NAColumn* x = ((BaseColumn *) ck)->getNAColumn();
if (((*x).*fptr)())
computedColumns += ck->getValueId();
}
return computedColumns;
}
void HbaseSearchSpec::addColumnNames(const ValueIdSet& vs)
{
// TEMP TEMP. Not all needed column names are being set up.
// for now, return without populating result.
// that will cause all columns to be retrieved.
//return;
for (ValueId vid = vs.init(); vs.next(vid); vs.advance(vid)) {
ItemExpr* ie = vid.getItemExpr();
NAString colName;
if ( ie->getOperatorType() == ITM_BASECOLUMN ) {
colName = ((BaseColumn*)ie)->getColName();
} else
if ( ie->getOperatorType() == ITM_INDEXCOLUMN ) {
colName = ((IndexColumn*)ie)->getNAColumn()->getIndexColName();
}
if (NOT colNames_.contains(colName))
colNames_.insert(colName);
}
}
// is any literal in this expr safely coercible to its target type?
NABoolean Cast::isSafelyCoercible(CacheWA &cwa) const
{
if (cwa.getPhase() >= CmpMain::BIND) {
ItemExpr *opd = child(0);
if (!opd->isSafelyCoercible(cwa)) {
return FALSE;
}
if (opd->getOperatorType() == ITM_CONSTANT) {
return ((ConstValue*)opd)->canBeSafelyCoercedTo(*type_);
}
return TRUE;
}
return FALSE;
}
// is val a constant that can be safely coerced to BaseColumn's type?
NABoolean BaseColumn::canSafelyCoerce(ItemExpr& val) const
{
NAColumn *baseCol = getNAColumn();
OperatorTypeEnum valTyp = val.getOperatorType();
return
(baseCol != NULL &&
((valTyp == ITM_CONSTANT &&
((ConstValue*)&val)->canBeSafelyCoercedTo(*baseCol->getType()))
// we don't need to check the types of param & hostvar here because we
// are called after ItemExpr::bindNode has successfully type-checked
// the current search predicate that is being tested for cacheability
|| valTyp == ITM_HOSTVAR
|| valTyp == ITM_DYN_PARAM
|| valTyp == ITM_CACHE_PARAM
));
}
// does this query's selection predicate list qualify query
// to be cacheable after this phase?
NABoolean ItemList::isListOfCacheableSelPred
(CacheWA& cwa, ItemList *other) const
{
Int32 arity = getArity();
NABoolean result = FALSE;
if (cwa.getPhase() >= CmpMain::BIND &&
other && arity == other->getArity()) {
// assume this is an AND list, so, we need only one
// cacheable conjunct to consider the list cacheable.
for (Int32 x = 0; x < arity; x++) {
ItemExpr *leftC = child(x), *rightC = other->child(x);
OperatorTypeEnum leftO = leftC->getOperatorType();
OperatorTypeEnum rightO = rightC->getOperatorType();
BaseColumn *base;
if (leftO == ITM_BASECOLUMN) {
base = (BaseColumn*)leftC;
if (base->isKeyColumnValue(*rightC)) {
cwa.addToUsedKeys(base);
}
result = TRUE;
continue;
}
else if (rightO == ITM_BASECOLUMN) {
base = (BaseColumn*)rightC;
if (base->isKeyColumnValue(*leftC)) {
cwa.addToUsedKeys(base);
}
result = TRUE;
continue;
}
else if (leftO == ITM_ITEM_LIST && rightO == ITM_ITEM_LIST &&
((ItemList*)leftC)->isListOfCacheableSelPred
(cwa, (ItemList*)rightC)) {
result = TRUE;
continue;
}
}
}
return result;
}
// is any literal in this expr safely coercible to its target type?
NABoolean Assign::isSafelyCoercible(CacheWA &cwa) const
{
if (cwa.getPhase() >= CmpMain::BIND) {
// we have to disallow caching of the following types of updates:
// update t set col = overlylongliteral
ItemExpr *src = getSource().getItemExpr();
if (src->getOperatorType() == ITM_CONSTANT) {
// source is a literal; make sure this update is cacheable only if
// literal can be safely coerced into its target type.
return ((ConstValue*)src)->canBeSafelyCoercedTo(getTarget().getType());
}
else { // source is not a literal
// we need to descend into this expr to verify that no
// errors can occur during backpatching. For example,
// update t set i = i + 123456789012345678901234567890
// should not be cacheable if i is a smallint.
// reject "update t set c=(subquery)" as noncacheable
// as part of a fix to CR 10-020108-8401.
return src->isSafelyCoercible(cwa) && src->isCacheableExpr(cwa);
}
}
return FALSE;
}
// check whether an element is in the collection
NABoolean ItemExprTreeAsList::contains(const ItemExpr *treeToCheck)
{
ItemExpr *aNodePtr = *treePtr_;
while (aNodePtr != NULL)
{
if (aNodePtr == treeToCheck)
return TRUE;
if (aNodePtr->getOperatorType() == operator_ AND
aNodePtr->getArity() >= 2)
{
if (shape_ == RIGHT_LINEAR_TREE)
aNodePtr = aNodePtr->child(1);
else
ABORT("can't do other than right-linear trees");
}
else
aNodePtr = NULL;
}
return FALSE;
}
// return number of entries
Lng32 ItemExprTreeAsList::entries() const
{
ItemExpr *aNode = *treePtr_;
Lng32 result = 0;
while (aNode != NULL)
{
result++;
if (aNode->getOperatorType() == operator_ AND
aNode->getArity() >= 2)
{
if (shape_ == LEFT_LINEAR_TREE)
aNode = aNode->child(0);
else
if (shape_ == RIGHT_LINEAR_TREE)
aNode = aNode->child(1);
else
ABORT("can't do other than right-linear trees");
}
else
aNode = NULL;
}
return result;
}
// change literals of a cacheable query into ConstantParameters
ItemExpr* UDFunction::normalizeForCache(CacheWA& cwa, BindWA& bindWA)
{
if (nodeIsNormalizedForCache()) {
return this;
}
// Since the UDFunction::transformNode refers to both the
// inputVars_ ValueIdSet and the children array we need to make sure
// they are consistent.
// The children array may contain cast() of the expressions in inputVars_
// If we have a UUDF function the inputs that were given
// at parse time, really do not reflect reality anymore.
// So here we will simply reinitialize them with what we find in the
// inputVars_.
CMPASSERT(udfDesc_); // we better have one after binding.
NABoolean isUUDF(udfDesc_->isUUDFRoutine());
// Save off a copy of the original inputVars_ set.
ValueIdSet origInputVars(inputVars_);
// Loop through the given inputs
for (Int32 x = 0; x < getArity(); x++)
{
NABoolean origInputIsChildOfCast(FALSE);
ItemExpr * origIe = child(x);
ItemExpr * newIe = NULL;
ValueId vid = origIe->getValueId();
if (cwa.getPhase() == CmpMain::BIND)
{
if (origIe->isSafelyCoercible(cwa))
{
// fix CR 10-010726-4109: make sure queries with constants that
// cannot be safely backpatched such as
// select case smallintcol when 4294967393 then 'max' end from t
// are not parameterized
newIe = origIe->normalizeForCache(cwa, bindWA);
if (newIe != origIe )
{
// normalizeForCache returned a new ItemExpr. We have to update
// the UDFunction inputVars_ set, as this is used to determine
// characteristic inputs for IsolatedScalarUDF at transform time.
child(x) = newIe;
// Is it a input that UDFunction::bindNode put a cast around?
if ((origIe->getOperatorType() == ITM_CAST) &&
(origInputVars.contains(origIe->child(0)->getValueId())))
{
// Since the original child was a CAST that UDFunction::bindNode
// created, check to see if that CAST's child is part of the new
// expression, if it is, we don't have to do anything, since the
// CASTED value is part of the inputVars set, otherwise, we have
// to update the inputVars set to keep the characteristic inputs
// consistent.
if (! newIe->referencesTheGivenValue(
origIe->child(0)->getValueId()), TRUE)
{
// remove the original input from inputVars. It is the child
// of origIe because origIe is a cast that UDFunction::bindNode
// introduced.
inputVars_ -= origIe->child(0)->getValueId();
if (newIe->getOperatorType() == ITM_CAST)
{
// If the new expression is a CAST, we assume the child is the
// real input. We don't expect CAST(CAST(CAST(expr))) type
// expressions only simple CAST(expr) ones.
inputVars_ += newIe->child(0)->getValueId();
}
else
{
// add the newIe itself if it was not a CAST.
inputVars_ += newIe->getValueId();
}
}
}
else
{
// If the newIe doesn't contain the original one, we need to update
// the inputVars set.
if (! newIe->referencesTheGivenValue(
origIe->getValueId()), TRUE)
{
// the origIe was not a CAST introduced by UDFunction::bindNode()
if (newIe->getOperatorType() == ITM_CAST)
{
if (!origInputVars.contains(newIe->child(0)->getValueId()))
{
inputVars_ -= origIe->getValueId();
// If the new expression is a CAST, we assume the child is the
// real input. We don't expect CAST(CAST(CAST(expr))) type
// expressions only simple CAST(expr) ones.
inputVars_ += newIe->child(0)->getValueId();
}
// we don't need to update inputVars_ if the origInputVars
// contains the valueId of the newIe child already.
}
else
{
// This is an entirely new input. Remove the old one, and
// add in the new.
inputVars_ -= origIe->getValueId();
inputVars_ += newIe->getValueId();
}
}
}
}
}
}
}
markAsNormalizedForCache();
return this;
}
// compress the histograms based on query predicates on this table
void TableDesc::compressHistogramsForCurrentQuery()
{
// if there are some column statistics
if ((colStats_.entries() != 0) &&
(table_) &&
(table_->getExtendedQualName().getSpecialType() == ExtendedQualName::NORMAL_TABLE))
{ // if 1
// check if query analysis info is available
if(QueryAnalysis::Instance()->isAnalysisON())
{ // if 2
// get a handle to the query analysis
QueryAnalysis* queryAnalysis = QueryAnalysis::Instance();
// get a handle to the table analysis
const TableAnalysis * tableAnalysis = getTableAnalysis();
if(!tableAnalysis)
return;
// iterate over statistics for each column
for(CollIndex i = 0; i < colStats_.entries(); i++)
{ // for 1
// Get a handle to the column's statistics descriptor
ColStatDescSharedPtr columnStatDesc = colStats_[i];
// get a handle to the ColStats
ColStatsSharedPtr colStats = columnStatDesc->getColStats();
// if this is a single column, as opposed to a multicolumn
if(colStats->getStatColumns().entries() == 1)
{ // if 3
// get column's value id
const ValueId columnId = columnStatDesc->getColumn();
// get column analysis
ColAnalysis* colAnalysis = queryAnalysis->getColAnalysis(columnId);
if(!colAnalysis) continue;
ValueIdSet predicatesOnColumn =
colAnalysis->getReferencingPreds();
// we can compress this column's histogram if there
// is a equality predicate against a constant
ItemExpr *constant = NULL;
NABoolean colHasEqualityAgainstConst =
colAnalysis->getConstValue(constant);
// if a equality predicate with a constant was found
// i.e. predicate of the form col = 5
if (colHasEqualityAgainstConst)
{ // if 4
if (constant)
// compress the histogram
columnStatDesc->compressColStatsForQueryPreds(constant,constant);
} // if 4
else { // else 4
// since there is no equality predicates we might still
// be able to compress the column's histogram based on
// range predicates against a constant. Following are
// examples of such predicates
// * col > 1 <-- predicate defines a lower bound
// * col < 3 <-- predicate defines a upper bound
// * col >1 and col < 30 <-- window predicate, define both bounds
ItemExpr * lowerBound = NULL;
ItemExpr * upperBound = NULL;
// Extract predicates from range spec and add it to the
// original predicate set otherwise isARangePredicate() will
// return FALSE, so histgram compression won't happen.
ValueIdSet rangeSpecPred(predicatesOnColumn);
for (ValueId predId= rangeSpecPred.init();
rangeSpecPred.next(predId);
rangeSpecPred.advance(predId))
{
ItemExpr * pred = predId.getItemExpr();
if ( pred->getOperatorType() == ITM_RANGE_SPEC_FUNC )
{
ValueIdSet vs;
((RangeSpecRef *)pred)->getValueIdSetForReconsItemExpr(vs);
// remove rangespec vid from the original set
predicatesOnColumn.remove(predId);
// add preds extracted from rangespec to the original set
predicatesOnColumn.insert(vs);
}
}
// in the following loop we iterate over all the predicates
// on this column. If there is a range predicate e.g. a > 2
// or a < 3, then we use that to define upper and lower bounds.
// Given predicate a > 2, we get a lower bound of 2.
// Given predicate a < 3, we get a upper bound of 3.
// The bound are then passed down to the histogram
// compression methods.
// iterate over predicates to see if any of them is a range
// predicate e.g. a > 2
for (ValueId predId= predicatesOnColumn.init();
predicatesOnColumn.next(predId);
predicatesOnColumn.advance(predId))
{ // for 2
// check if this predicate is a range predicate
ItemExpr * predicateOnColumn = predId.getItemExpr();
if (predicateOnColumn->isARangePredicate())
{ // if 5
// if a predicate is a range predicate we need to find out more
// information regarding the predicate to see if it can be used
// to compress the columns histogram. We look for the following:
// * The predicate is against a constant e.g. a > 3 and not against
// another column e.g. a > b
// Also give a predicate we need to find out what side is the column
// and what side is the constant. Normally people write a range predicate
// as a > 3, but the same could be written as 3 < a.
// Also either on of the operands of the range predicate might be
// a VEG, if so then we need to dig into the VEG to see where is
// the constant and where is the column.
// check the right and left children of this predicate to
// see if one of them is a constant
ItemExpr * leftChildItemExpr = (ItemExpr *) predicateOnColumn->getChild(0);
ItemExpr * rightChildItemExpr = (ItemExpr *) predicateOnColumn->getChild(1);
// by default assume the literal is at right i.e. predicate of
// the form a > 2
NABoolean columnAtRight = FALSE;
// check if right child of predicate is a VEG
if ( rightChildItemExpr->getOperatorType() == ITM_VEG_REFERENCE)
{ // if 6
// if child is a VEG
VEGReference * rightChildVEG = (VEGReference *) rightChildItemExpr;
// check if the VEG contains the current column
// if it does contain the current column then
// the predicate has the column on right and potentially
// a constant on the left.
if(rightChildVEG->getVEG()->getAllValues().contains(columnId))
{ // if 7
// column is at right i.e. predicate is of the form
// 2 < a
columnAtRight = TRUE;
} // if 7
} // if 6
else { // else 6
// child is not a VEG
if ( columnId == rightChildItemExpr->getValueId() )
{ // if 8
// literals are at left i.e. predicate is of the form
// (1,2) < (a, b)
columnAtRight = TRUE;
} // if 8
} // else 6
ItemExpr * potentialConstantExpr = NULL;
// check if the range predicate is against a constant
if (columnAtRight)
{ // if 9
// the left child is potentially a constant
potentialConstantExpr = leftChildItemExpr;
} // if 9
else { // else 9
// the right child is potentially a constant
potentialConstantExpr = rightChildItemExpr;
} // else 9
// initialize constant to NULL before
// looking for next constant
constant = NULL;
// check if potentialConstantExpr contains a constant.
// we need to see if this range predicate is a predicate
// against a constant e.g col > 1 and not a predicate
// against another column e.g. col > anothercol
// if the expression is a VEG
if ( potentialConstantExpr->getOperatorType() == ITM_VEG_REFERENCE)
{ // if 10
// expression is a VEG, dig into the VEG to
// get see if it contains a constant
VEGReference * potentialConstantExprVEG =
(VEGReference *) potentialConstantExpr;
potentialConstantExprVEG->getVEG()->\
getAllValues().referencesAConstValue(&constant);
} // if 10
else { // else 10
// express is not a VEG, it is a constant
if ( potentialConstantExpr->getOperatorType() == ITM_CONSTANT )
constant = potentialConstantExpr;
} // else 10
// if predicate involves a constant, does the constant imply
// a upper bound or lower bound
if (constant)
{ // if 11
// if range predicate has column at right e.g. 3 > a
if (columnAtRight)
{ // if 12
if ( predicateOnColumn->getOperatorType() == ITM_GREATER ||
predicateOnColumn->getOperatorType() == ITM_GREATER_EQ)
{ // if 13
if (!upperBound)
upperBound = constant;
} // if 13
else
{ // else 13
if (!lowerBound)
lowerBound = constant;
} // else 13
} // if 12
else { // else 12
// range predicate has column at left e.g. a < 3
if ( predicateOnColumn->getOperatorType() == ITM_LESS ||
predicateOnColumn->getOperatorType() == ITM_LESS_EQ)
{ // if 14
if (!upperBound)
upperBound = constant;
} // if 14
else
{ // else 14
if (!lowerBound)
lowerBound = constant;
} // else 14
} // else 12
} // if 11
} // if 5
} // for 2
// if we found a upper bound or a lower bound
if (lowerBound || upperBound)
{
// compress the histogram based on range predicates
columnStatDesc->compressColStatsForQueryPreds(lowerBound, upperBound);
}
} // else 4
} // if 3
} // for 1
} // if 2
} // if 1
// All histograms compressed. Set the histCompressed flag to TRUE
histsCompressed(TRUE);
}
// 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
// insert a new entry.
// The new entry is created at the end of the current tree.
void ItemExprTreeAsList::insert(ItemExpr *treeToInsert)
{
if (treePtr_->getPtr() == NULL)
{
*treePtr_ = treeToInsert;
}
else
{
if (shape_ == RIGHT_LINEAR_TREE)
{
switch (operator_)
{
case ITM_AND:
*treePtr_ = new(CmpCommon::statementHeap())
BiLogic(operator_,
treeToInsert,
treePtr_->getPtr());
break;
case ITM_ITEM_LIST:
{
ItemExpr *aNode = treePtr_->getPtr();
ItemExpr *pNode = treePtr_->getPtr();
if (aNode->getOperatorType() != operator_ AND
aNode->getArity() < 2)
{
// case of current number of entries equal to 1.
*treePtr_ = new(CmpCommon::statementHeap())
ItemList(treePtr_->getPtr(),
treeToInsert);
}
else
{
// current number of entries > 1
while (aNode != NULL)
{
if (aNode->getOperatorType() == operator_ AND
aNode->getArity() >= 2)
{
if (shape_ == RIGHT_LINEAR_TREE)
{
pNode = aNode;
aNode = aNode->child(1);
}
else
ABORT("can't do other than right-linear trees");
}
else
aNode = NULL;
}
pNode->child(1) = new(CmpCommon::statementHeap())
ItemList(pNode->child(1),
treeToInsert);
}
}
break;
default:
ABORT("Can't do backbones with this node type");
}
}
else
ABORT("can only insert into right-linear trees");
}
}
// 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
// remove an element that is given by its value
ItemExpr * ItemExprTreeAsList::remove(ItemExpr *treeToRemove)
{
ItemExpr *andNodePtr = treePtr_->getPtr();
ItemExpr *predecessor = NULL;
NABoolean found = FALSE;
// assume the predicate is represented in right-linear
// form:
//
// op
// / \
// A op
// / \
// B ...
//
// and search for the <op> node that is directly above the
// predicate that we want to remove.
if (shape_ != RIGHT_LINEAR_TREE)
ABORT("delete is supported for right linear trees only");
// is the node to delete the only node?
if (treeToRemove == andNodePtr)
{
found = TRUE;
delete andNodePtr;
*treePtr_ = NULL;
}
else
// traverse the backbone looking for "treeToRemove"
while (andNodePtr != NULL AND
andNodePtr->getOperatorType() == operator_ AND
andNodePtr->getArity() == 2 AND
NOT found)
{
// did we find the right node?
if (andNodePtr->child(0).getPtr() == treeToRemove)
{
found = TRUE;
// take "andNode" out of the original tree
if (predecessor != NULL)
{
predecessor->child(1) = andNodePtr->child(1);
}
else
*treePtr_ = andNodePtr->child(1);
// set all children of "andNode" to NULL, then delete it
andNodePtr->child(0) = NULL;
andNodePtr->child(1) = NULL;
delete andNodePtr;
andNodePtr = NULL;
}
else
{
predecessor = andNodePtr;
andNodePtr = andNodePtr->child(1);
}
}
if (found)
return treeToRemove;
else
return NULL;
}
// -----------------------------------------------------------------------
// make an IndexDesc from an existing TableDesc and an NAFileSet
// -----------------------------------------------------------------------
IndexDesc::IndexDesc(TableDesc *tdesc,
NAFileSet *fileSet,
CmpContext* cmpContext)
: tableDesc_(tdesc), clusteringIndexFlag_(FALSE),
identityColumnUniqueIndexFlag_(FALSE), partFunc_(NULL),
fileSet_(fileSet), cmpContext_(cmpContext), scanBasicCosts_(NULL)
{
DCMPASSERT( tdesc != NULL AND fileSet != NULL );
Lng32 ixColNumber;
ValueId keyValueId;
ValueId baseValueId;
const NATable *naTable = tdesc->getNATable();
indexLevels_ = fileSet_->getIndexLevels();
// ---------------------------------------------------------------------
// Make the column list for the index or vertical partition.
// Any reference to index also holds for vertical partitions.
// ---------------------------------------------------------------------
const NAColumnArray & allColumns = fileSet_->getAllColumns();
// any index gets a new set of IndexColumn
// item expressions and new value ids
CollIndex i = 0;
for (i = 0; i < allColumns.entries(); i++)
{
ItemExpr *baseItemExpr = NULL;
// make a new IndexColumn item expression, indicate how it is
// defined (in terms of base table columns) and give a value
// id to the new IndexColumn expression
if (allColumns[i]->getPosition() >= 0)
{
baseValueId =
tdesc->getColumnList()[allColumns[i]->getPosition()];
baseItemExpr = baseValueId.getItemExpr();
}
else
{
// this column doesn't exist in the base table.
// This is the KEYTAG column of sql/mp indices.
ItemExpr * keytag = new(wHeap())
NATypeToItem((NAType *)(allColumns[i]->getType()));
keytag->synthTypeAndValueId();
baseValueId = keytag->getValueId();
baseItemExpr = NULL;
}
#pragma nowarn(1506) // warning elimination
IndexColumn *ixcol = new(wHeap()) IndexColumn(fileSet_,i,baseValueId);
#pragma warn(1506) // warning elimination
ixcol->synthTypeAndValueId();
// add the newly obtained value id to the index column list
indexColumns_.insert(ixcol->getValueId());
// if the index column is defined as a 1:1 copy of a base
// column, add it as an equivalent index column (EIC) to the
// base column item expression
if ((baseItemExpr) &&
(baseItemExpr->getOperatorType() == ITM_BASECOLUMN))
((BaseColumn *) baseItemExpr)->addEIC(ixcol->getValueId());
}
// ---------------------------------------------------------------------
// make the list of access key columns in the index and make a list
// of the order that the index provides
// ---------------------------------------------------------------------
const NAColumnArray & indexKeyColumns = fileSet_->getIndexKeyColumns();
for (i = 0; i < indexKeyColumns.entries(); i++)
{
// which column of the index is this (usually this will be == i)
#pragma nowarn(1506) // warning elimination
if ( !naTable->isHbaseTable() )
ixColNumber = allColumns.index(indexKeyColumns[i]);
else {
// For Hbase tables, a new NAColumn is created for every column
// in an index. The above pointer-based lookup for the key column
// in base table will only find the index column itself. The
// fix is to lookup by the column name and type as is
// implemented by the getColumnPosition() method.
ixColNumber = allColumns.getColumnPosition(*indexKeyColumns[i]);
CMPASSERT(ixColNumber >= 0);
}
#pragma warn(1506) // warning elimination
// insert the value id of the index column into the key column
// value id list
keyValueId = indexColumns_[ixColNumber];
indexKey_.insert(keyValueId);
// insert the same value id into the order list, if the column
// is in ascending order, otherwise insert the inverse of the
// column
if (indexKeyColumns.isAscending(i))
{
orderOfKeyValues_.insert(keyValueId);
}
else
{
InverseOrder *invExpr = new(wHeap())
InverseOrder(keyValueId.getItemExpr());
invExpr->synthTypeAndValueId();
orderOfKeyValues_.insert(invExpr->getValueId());
}
}
markIdentityColumnUniqueIndex(tdesc);
// ---------------------------------------------------------------------
// Find the clustering key columns in the index and store their value
// ids in clusteringKey_
// ---------------------------------------------------------------------
NABoolean found = TRUE;
const NAColumnArray & clustKeyColumns =
naTable->getClusteringIndex()->getIndexKeyColumns();
for (i = 0; i < clustKeyColumns.entries() AND found; i++)
{
// which column of the index is this?
#pragma nowarn(1506) // warning elimination
ixColNumber = allColumns.index(clustKeyColumns[i]);
#pragma warn(1506) // warning elimination
found = (ixColNumber != NULL_COLL_INDEX);
if (found)
{
// insert the value id of the index column into the clustering key
// value id list
keyValueId = indexColumns_[ixColNumber];
clusteringKey_.insert(keyValueId);
}
else
{
// clustering key isn't contained in this index, clear the
// list that is supposed to indicate the clustering key
clusteringKey_.clear();
}
}
// ---------------------------------------------------------------------
// make the list of partitioning key columns in the index and make a list
// of the order that the partitioning provides
// ---------------------------------------------------------------------
const NAColumnArray & partitioningKeyColumns
= fileSet_->getPartitioningKeyColumns();
for (i = 0; i < partitioningKeyColumns.entries(); i++)
{
// which column of the index is this
#pragma nowarn(1506) // warning elimination
ixColNumber = allColumns.index(partitioningKeyColumns[i]);
#pragma warn(1506) // warning elimination
// insert the value id of the index column into the partitioningkey column
// value id list
keyValueId = indexColumns_[ixColNumber];
partitioningKey_.insert(keyValueId);
// insert the same value id into the order list, if the column
// is in ascending order, otherwise insert the inverse of the
// column
if (partitioningKeyColumns.isAscending(i))
{
orderOfPartitioningKeyValues_.insert(keyValueId);
}
else
{
InverseOrder *invExpr = new(wHeap())
InverseOrder(keyValueId.getItemExpr());
invExpr->synthTypeAndValueId();
orderOfPartitioningKeyValues_.insert(invExpr->getValueId());
}
}
// ---------------------------------------------------------------------
// If this index is partitioned, find the partitioning key columns
// and build a partitioning function.
// ---------------------------------------------------------------------
if ((fileSet_->getCountOfFiles() > 1) ||
(fileSet_->getPartitioningFunction() &&
fileSet_->getPartitioningFunction()->
isARoundRobinPartitioningFunction()))
partFunc_ = fileSet_->getPartitioningFunction()->
createPartitioningFunctionForIndexDesc(this);
} // IndexDesc::IndexDesc()
// ItmSeqRunningFunction::preCodeGen
//
// Transforms the running sequence functions into scalar expressions
// that use offset to reference the previous value of the function.
//
ItemExpr *ItmSeqRunningFunction::preCodeGen(Generator *generator)
{
if (nodeIsPreCodeGenned())
return this;
markAsPreCodeGenned();
// Get some local handles...
//
CollHeap *wHeap = generator->wHeap();
ItemExpr *itmChild = child(0)->castToItemExpr();
// Allocate a HostVar for referencing the result before it is computed.
//
ItemExpr *itmResult
= new(wHeap) HostVar("_sys_Result",
getValueId().getType().newCopy(wHeap),
TRUE);
// Create an item expression to reference the previous
// value of this running sequence function.
//
ItemExpr *offExpr = new(wHeap) ItmSeqOffset(itmResult, 1);
((ItmSeqOffset *)offExpr)->setIsOLAP(isOLAP());
// Add the sequence function specific computation.
//
ItemExpr *itmNewSeqFunc = 0;
switch(getOperatorType())
{
case ITM_RUNNING_COUNT:
{
// By this point ITM_RUNNING_COUNT is count(column). The count
// is one more than the previous count if the current column is
// not null, otherwise, it is the previous count.
//
// Create the increment value. For non-nullable values, this
// is always 1, essentially runningcount(*).
//
ItemExpr *incr;
if(itmChild->getValueId().getType().supportsSQLnullLogical()) {
incr = generator->getExpGenerator()->createExprTree
("CASE WHEN @A1 IS NULL THEN @A3 ELSE @A2 END",
0, 3,
itmChild,
new(wHeap) ConstValue(1),
new(wHeap) ConstValue(0));
} else {
incr = new(wHeap) ConstValue(1);
}
((ItmSeqOffset *)offExpr)->setNullRowIsZero(TRUE);
ItemExpr *src = offExpr;
// Do the increment.
//
itmNewSeqFunc = new(wHeap)
BiArith(ITM_PLUS, src, incr);
}
break;
case ITM_RUNNING_SUM:
{
// SUM(sum from previous row, child)
//
itmNewSeqFunc = new(wHeap) BiArithSum(ITM_PLUS, offExpr, itmChild);
}
break;
case ITM_RUNNING_MIN:
{
// MIN(min from previous rows, child)
//
itmNewSeqFunc
= new(wHeap) ItmScalarMinMax(ITM_SCALAR_MIN,
offExpr,
itmChild);
}
break;
case ITM_RUNNING_MAX:
{
// MAX(max from previous row, child)
//
itmNewSeqFunc
= new(wHeap) ItmScalarMinMax(ITM_SCALAR_MAX,
offExpr,
itmChild);
}
break;
case ITM_LAST_NOT_NULL:
{
// If the current value is null then use the previous value
// of last not null.
//
itmNewSeqFunc = generator->getExpGenerator()->createExprTree
("CASE WHEN @A2 IS NOT NULL THEN @A2 ELSE @A1 END",
0, 2, offExpr, itmChild);
}
break;
case ITM_RUNNING_CHANGE:
{
// The running change (or 'rows since changed') can have a
// composite child (a list of values)
// Convert the list of values to a list of offset of values.
//
ItemExpr *offChild = itmChild;
if (itmChild->getOperatorType() == ITM_ITEM_LIST)
{
// child is a multi-valued expression, transform into multiple
//
ExprValueId treePtr = itmChild;
ItemExprTreeAsList changeValues(&treePtr,
ITM_ITEM_LIST,
RIGHT_LINEAR_TREE);
offChild = new(wHeap) ItmSeqOffset( changeValues[0], 1);
((ItmSeqOffset *)offChild)->setIsOLAP(isOLAP());
// add Offset expressions for all the items of the list
//
CollIndex nc = changeValues.entries();
for (CollIndex i = 1; i < nc; i++)
{
ItemExpr *off = new(generator->wHeap()) ItmSeqOffset( changeValues[i], 1);
((ItmSeqOffset *)off)->setIsOLAP(isOLAP());
offChild = new(generator->wHeap()) ItemList(offChild, off);
}
} else {
offChild = new(wHeap) ItmSeqOffset( offChild, 1);
((ItmSeqOffset *)offChild)->setIsOLAP(isOLAP());
}
((ItmSeqOffset *)offExpr)->setNullRowIsZero(TRUE);
ItemExpr *prevValue = offExpr;
// Compare the value(s) to the previous value(s). Use special
// NULLs flags to treat NULLs as values. Two NULL values are
// considered equal here.
//
ItemExpr *pred = new (wHeap) BiRelat(ITM_EQUAL,
itmChild,
offChild,
TRUE); // Special NULLs
// running change =
// (value(s) == prev(value(s))) ? prev(running change)+1 : 1
//
// Compute base value.
//
itmNewSeqFunc = new (wHeap)
IfThenElse(pred, prevValue, new (wHeap) SystemLiteral(0));
itmNewSeqFunc = new (wHeap) Case(NULL, itmNewSeqFunc);
// Force the evaluation of the offset expression so that the
// result can be reused by subsequent references.
//
itmNewSeqFunc = new(wHeap) ItmBlockFunction(offChild, itmNewSeqFunc);
// Increment the base value.
//
itmNewSeqFunc = new (wHeap) BiArith(ITM_PLUS,
itmNewSeqFunc,
new(wHeap) SystemLiteral(1));
}
break;
}
// Get value Ids and types for all of the items. Must do this typing before
// replacing this value Id's item expression -- otherwise, the typing
// will give a result different than the type already computed for this
// sequence function.
//
GenAssert(itmNewSeqFunc, "ItmSeqRunningFunction::preCodeGen -- Unexpected Operator Type!");
itmNewSeqFunc->synthTypeAndValueId(TRUE);
// Replace the original value ID with the new expression.
//
getValueId().replaceItemExpr(itmNewSeqFunc);
// Map the reference to the result to the actual result in the map table.
//
Attributes *attr = generator->getMapInfo
(itmNewSeqFunc->getValueId())->getAttr();
MapInfo *mapInfo = generator->addMapInfo(itmResult->getValueId(), attr);
itmResult->markAsPreCodeGenned();
mapInfo->codeGenerated();
// Return the preCodeGen of the new expression.
//
return itmNewSeqFunc->preCodeGen(generator);
}
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)
short BiLogic::codeGen(Generator * generator)
{
Attributes ** attr;
if (generator->getExpGenerator()->genItemExpr(this, &attr, (1+getArity()), 0) == 1)
return 0;
Space * space = generator->getSpace();
ExpGenerator * expGen = generator->getExpGenerator();
// Normally, if code for a value id has been generated, and if
// that value id is seen again, then code is not generated. The
// location where the result is available is returned instead.
// The case of a logical operator is different. Code is generated
// again if a value id from the left child is also present in
// the right child. This is done
// because at expression evaluation time, some of the expressions
// may be skipped due to short circuit evaluation.
//
// Allocate a new map table before generating code for each child.
// This map table contains all the temporary results produced by
// the child.
// Remove this map table after generating code for each child.
generator->appendAtEnd();
expGen->incrementLevel();
codegen_and_set_attributes(generator, attr, 2);
// generator->getExpGenerator()->setClauseLinked(FALSE);
// child(0)->codeGen(generator);
// attr[1] = generator->getAttr(child(0));
// generator->getExpGenerator()->setClauseLinked(FALSE);
/* generate boolean short circuit code */
Attributes ** branch_attr = new(generator->wHeap()) Attributes * [2];
branch_attr[0] = attr[0]->newCopy(generator->wHeap());
branch_attr[0]->copyLocationAttrs(attr[0]);
branch_attr[1] = attr[1]->newCopy(generator->wHeap());
branch_attr[1]->copyLocationAttrs(attr[1]);
branch_attr[0]->resetShowplan();
ex_branch_clause * branch_clause
= new(space) ex_branch_clause(getOperatorType(), branch_attr, space);
generator->getExpGenerator()->linkClause(0, branch_clause);
generator->removeLast();
expGen->decrementLevel();
generator->appendAtEnd();
expGen->incrementLevel();
ValueIdSet markedEntries;
// This ia a MapTable entry related fix for RangeSpec transformation.
if( child(1)->getOperatorType() == ITM_RANGE_SPEC_FUNC )
markGeneratedEntries(generator, child(1)->child(1), markedEntries);
else
markGeneratedEntries(generator, child(1), markedEntries);
// if( child(1)->getOperatorType() == ITM_RANGE_SPEC_FUNC )
// child(1)->child(1)->codeGen(generator);
// else
child(1)->codeGen(generator);
ItemExpr *rightMost;
if( child(1)->getOperatorType() == ITM_RANGE_SPEC_FUNC )
rightMost = child(1)->child(1)->castToItemExpr();
else
rightMost = child(1)->castToItemExpr();
while (rightMost->getOperatorType() == ITM_ITEM_LIST)
rightMost = rightMost->child(1)->castToItemExpr();
attr[2] = generator->
getMapInfo(rightMost->getValueId())->getAttr();
ex_bool_clause * bool_clause =
new(space) ex_bool_clause(getOperatorType(), attr, space);
generator->getExpGenerator()->linkClause(this, bool_clause);
branch_clause->set_branch_clause((ex_clause *)bool_clause);
generator->removeLast();
expGen->decrementLevel();
if( child(1)->getOperatorType() == ITM_RANGE_SPEC_FUNC )
unGenerate(generator, child(1)->child(1));
else
unGenerate(generator, child(1));
generateMarkedEntries(generator, markedEntries);
return 0;
}
//----------------------------------------------------------------------------
// Transform each aggregate in the MAV select list, to use the @OP column in
// order to calculate deleted rows correctly:
// COUNT(*) => SUM(@OP)
// COUNT(a) => SUM(IF (a IS NULL) THEN 0 ELSE @OP)
// SUM(a) => SUM(a * @OP)
// MIN(a) & MAX(a) => Special treatment (see handleDeltaMinMaxColumns()).
// AVG, STDDEV and VARIANCE are derived from other COUNT and SUM columns.
// They are removed from this select list to avoid divide by zero problems
// in case COUNT=0.
void MavRelRootBuilder::fixDeltaColumns(RelExpr *mvSelectTree,
NABoolean canSkipMinMax,
NABoolean wasFullDE)
{
// Get the MAV select list from the mvSelectTree.
CMPASSERT(mvSelectTree->getOperatorType() == REL_ROOT);
RelRoot *mavSelectRoot = (RelRoot *)mvSelectTree;
ItemExprList mavSelectList(mavSelectRoot->removeCompExprTree(), heap_);
ItemExprList newSelectList(heap_);
const NAString& opName = MavBuilder::getVirtualOpColumnName();
for (CollIndex i=0; i<mavSelectList.entries(); i++)
{
ItemExpr *selectListExpr = mavSelectList[i];
CMPASSERT(selectListExpr->getOperatorType() == ITM_RENAME_COL);
const NAString& colName =
((RenameCol*)selectListExpr)->getNewColRefName()->getColName();
ItemExpr *aggregateExpr = selectListExpr;
while ((aggregateExpr->getOperatorType() == ITM_RENAME_COL) ||
(aggregateExpr->getOperatorType() == ITM_CAST))
aggregateExpr = aggregateExpr->child(0);
// Assuming all aggregates are top-most functions as explained in the
// MV external spec.
ItemExpr *aggregateOperand = aggregateExpr->child(0);
ItemExpr *newExpr = NULL;
if (!aggregateExpr->containsAnAggregate())
{
// Non-aggregate columns (group-by cols) are added unchanged.
}
else
switch (aggregateExpr->getOperatorType())
{
case ITM_COUNT_NONULL:
// COUNT(a) => SUM(IF (a IS NULL) THEN 0 ELSE @OP)
newExpr = new(heap_)
Aggregate(ITM_SUM, new(heap_)
Case(NULL, new(heap_)
IfThenElse (new(heap_) UnLogic(ITM_IS_NULL, aggregateOperand),
new(heap_) SystemLiteral(0),
new(heap_) ColReference(new(heap_)
ColRefName(opName)))));
selectListExpr->child(0) = newExpr;
break;
case ITM_COUNT:
// COUNT(*) => SUM(@OP)
newExpr = new(heap_)
Aggregate(ITM_SUM,
new(heap_) ColReference(new(heap_) ColRefName(opName)));
selectListExpr->child(0) = newExpr;
break;
case ITM_SUM:
// SUM(a) => SUM(a * @OP)
newExpr = new(heap_)
BiArith(ITM_TIMES,
aggregateOperand,
new(heap_) ColReference(new(heap_) ColRefName(opName)));
aggregateExpr->child(0) = newExpr;
break;
case ITM_MIN:
case ITM_MAX:
// Handle Min/Max if not all deltas are INSERT ONLY.
if (!canSkipMinMax)
handleDeltaMinMaxColumns(aggregateExpr, colName, newSelectList, wasFullDE);
// Add the SYS_DELTA Min/Max aggregate as is.
break;
case ITM_AVG:
case ITM_DIVIDE:
case ITM_VARIANCE:
case ITM_STDDEV:
// Remove these expressions from the select list.
selectListExpr = NULL;
break;
default:
// A new type of aggregate?
CMPASSERT(FALSE);
break;
}
if (selectListExpr!=NULL)
newSelectList.insert(selectListExpr);
}
ItemExpr *selList = newSelectList.convertToItemExpr();
mavSelectRoot->addCompExprTree(selList);
mavSelectList.clear(); // Don't delete all the entries from the Dtor.
} // MavRelRootBuilder::fixDeltaColumns()