void MvQueryRewriteHandler::dumpAnalysisToFile(QueryAnalysis* qa, RelExpr* expr)
{
// Dump the QueryAnalysis data to a file.
NAString analysisFileName = fileNamePrefix_ + ".analysis";
NAString str;
expr->unparse(str, OPTIMIZER_PHASE, MVINFO_FORMAT);
str += "\n";
str += qa->getText();
// Add in some stuff to look at join predicates for the JBBCs.
str += "Join Predicates\n";
str += "===============";
char buffer[20];
ARRAY(JBB*) jbbs = qa->getJBBs();
for (CollIndex jbbInx = 0; jbbInx < jbbs.entries(); jbbInx++)
{
JBB* jbb = jbbs[jbbInx];
str_itoa(jbbInx, buffer);
((str += "\nJBB #") += NAString(buffer)) += ":\n";
CANodeIdSet jbbcs = jbb->getJBBCs();
for (CANodeId jbbcId=jbbcs.init(); jbbcs.next(jbbcId); jbbcs.advance(jbbcId) )
{
str_itoa(jbbcId, buffer);
((str += "\nJBBC with CANodeId ") += NAString(buffer)) += ":\n";
ValueIdSet joinPreds = jbbcId.getNodeAnalysis()->getJBBC()->getJoinPreds();
str += valueIdSetGetText(joinPreds);
if (joinPreds.entries() > 0)
{
str.append("\n(value ids of predicates are ");
NABoolean first = true;
for (ValueId jpVid=joinPreds.init(); joinPreds.next(jpVid); joinPreds.advance(jpVid))
{
if (first)
first = FALSE;
else
str.append(", ");
str_itoa(jpVid, buffer);
str.append(buffer);
}
str.append(")\n");
}
}
str += '\n';
}
dumpToFile(analysisFileName.data(), str.data());
} // dumpAnalysisToFile()
short
PhysSequence::codeGen(Generator *generator)
{
// Get a local handle on some of the generator objects.
//
CollHeap *wHeap = generator->wHeap();
Space *space = generator->getSpace();
ExpGenerator *expGen = generator->getExpGenerator();
MapTable *mapTable = generator->getMapTable();
// Allocate a new map table for this node. This must be done
// before generating the code for my child so that this local
// map table will be sandwiched between the map tables already
// generated and the map tables generated by my offspring.
//
// Only the items available as output from this node will
// be put in the local map table. Before exiting this function, all of
// my offsprings map tables will be removed. Thus, none of the outputs
// from nodes below this node will be visible to nodes above it except
// those placed in the local map table and those that already exist in
// my ancestors map tables. This is the standard mechanism used in the
// generator for managing the access to item expressions.
//
MapTable *localMapTable = generator->appendAtEnd();
// Since this operation doesn't modify the row on the way down the tree,
// go ahead and generate the child subtree. Capture the given composite row
// descriptor and the child's returned TDB and composite row descriptor.
//
ex_cri_desc * givenCriDesc = generator->getCriDesc(Generator::DOWN);
child(0)->codeGen(generator);
ComTdb *childTdb = (ComTdb*)generator->getGenObj();
ex_cri_desc * childCriDesc = generator->getCriDesc(Generator::UP);
ExplainTuple *childExplainTuple = generator->getExplainTuple();
// Make all of my child's outputs map to ATP 1. The child row is only
// accessed in the project expression and it will be the second ATP
// (ATP 1) passed to this expression.
//
localMapTable->setAllAtp(1);
// My returned composite row has an additional tupp.
//
Int32 numberTuples = givenCriDesc->noTuples() + 1;
ex_cri_desc * returnCriDesc
#pragma nowarn(1506) // warning elimination
= new (space) ex_cri_desc(numberTuples, space);
#pragma warn(1506) // warning elimination
// For now, the history buffer row looks just the return row. Later,
// it may be useful to add an additional tupp for sequence function
// itermediates that are not needed above this node -- thus, this
// ATP is kept separate from the returned ATP.
//
const Int32 historyAtp = 0;
const Int32 historyAtpIndex = numberTuples-1;
#pragma nowarn(1506) // warning elimination
ex_cri_desc *historyCriDesc = new (space) ex_cri_desc(numberTuples, space);
#pragma warn(1506) // warning elimination
ExpTupleDesc *historyDesc = 0;
//seperate the read and retur expressions
seperateReadAndReturnItems(wHeap);
// The history buffer consists of items projected directly from the
// child, the root sequence functions, the value arguments of the
// offset functions, and running sequence functions. These elements must
// be materialized in the history buffer in order to be able to compute
// the outputs of this node -- the items projected directly from the child
// (projectValues) and the root sequence functions (sequenceFunctions).
//
// Compute the set of sequence function items that must be materialized
// int the history buffer. -- sequenceItems
//
// Compute the set of items in the history buffer: the union of the
// projected values and the value arguments. -- historyIds
//
// Compute the set of items in the history buffer that are computed:
// the difference between all the elements in the history buffer
// and the projected items. -- computedHistoryIds
//
// KB---will need to return atp with 3 tups only 0,1 and 2
// 2 -->values from history buffer after ther are moved to it
addCheckPartitionChangeExpr(generator, TRUE);
ValueIdSet historyIds;
historyIds += movePartIdsExpr();
historyIds += sequencedColumns();
ValueIdSet outputFromChild = child(0)->getGroupAttr()->getCharacteristicOutputs();
getHistoryAttributes(readSeqFunctions(),outputFromChild, historyIds, TRUE, wHeap);
// Add in the top level sequence functions.
historyIds += readSeqFunctions();
getHistoryAttributes(returnSeqFunctions(),outputFromChild, historyIds, TRUE, wHeap);
// Add in the top level functions.
historyIds += returnSeqFunctions();
// Layout the work tuple format which consists of the projected
// columns and the computed sequence functions. First, compute
// the number of attributes in the tuple.
//
ULng32 numberAttributes
= ((NOT historyIds.isEmpty()) ? historyIds.entries() : 0);
// Allocate an attribute pointer vector from the working heap.
//
Attributes **attrs = new(wHeap) Attributes*[numberAttributes];
// Fill in the attributes vector for the history buffer including
// adding the entries to the map table. Also, compute the value ID
// set for the elements to project from the child row.
//
//??????????re-visit this function??
computeHistoryAttributes(generator,
localMapTable,
attrs,
historyIds);
// Create the tuple descriptor for the history buffer row and
// assign the offsets to the attributes. For now, this layout is
// identical to the returned row. Set the tuple descriptors for
// the return and history rows.
//
ULng32 historyRecLen;
expGen->processAttributes(numberAttributes,
attrs,
ExpTupleDesc::SQLARK_EXPLODED_FORMAT,
historyRecLen,
historyAtp,
historyAtpIndex,
&historyDesc,
ExpTupleDesc::SHORT_FORMAT);
NADELETEBASIC(attrs, wHeap);
#pragma nowarn(1506) // warning elimination
returnCriDesc->setTupleDescriptor(historyAtpIndex, historyDesc);
#pragma warn(1506) // warning elimination
#pragma nowarn(1506) // warning elimination
historyCriDesc->setTupleDescriptor(historyAtpIndex, historyDesc);
#pragma warn(1506) // warning elimination
// If there are any sequence function items, generate the sequence
// function expressions.
//
ex_expr * readSeqExpr = NULL;
if(NOT readSeqFunctions().isEmpty())
{
ValueIdSet seqVals = readSeqFunctions();
seqVals += sequencedColumns();
seqVals += movePartIdsExpr();
expGen->generateSequenceExpression(seqVals,
readSeqExpr);
}
ex_expr *checkPartChangeExpr = NULL;
if (!checkPartitionChangeExpr().isEmpty()) {
ItemExpr * newCheckPartitionChangeTree=
checkPartitionChangeExpr().rebuildExprTree(ITM_AND,TRUE,TRUE);
expGen->generateExpr(newCheckPartitionChangeTree->getValueId(),
ex_expr::exp_SCAN_PRED,
&checkPartChangeExpr);
}
//unsigned long rowLength;
ex_expr * returnExpr = NULL;
if(NOT returnSeqFunctions().isEmpty())
{
expGen->generateSequenceExpression(returnSeqFunctions(),
returnExpr);
}
// Generate expression to evaluate predicate on the output
//
ex_expr *postPred = 0;
if (! selectionPred().isEmpty()) {
ItemExpr * newPredTree =
selectionPred().rebuildExprTree(ITM_AND,TRUE,TRUE);
expGen->generateExpr(newPredTree->getValueId(), ex_expr::exp_SCAN_PRED,
&postPred);
}
// Reset ATP's to zero for parent.
//
localMapTable->setAllAtp(0);
// Generate expression to evaluate the cancel expression
//
ex_expr *cancelExpression = 0;
if (! cancelExpr().isEmpty()) {
ItemExpr * newCancelExprTree =
cancelExpr().rebuildExprTree(ITM_AND,TRUE,TRUE);
expGen->generateExpr(newCancelExprTree->getValueId(), ex_expr::exp_SCAN_PRED,
&cancelExpression);
}
//
// For overflow
//
// ( The following are meaningless if ! unlimitedHistoryRows() )
NABoolean noOverflow =
CmpCommon::getDefault(EXE_BMO_DISABLE_OVERFLOW) == DF_ON ;
NABoolean logDiagnostics =
CmpCommon::getDefault(EXE_DIAGNOSTIC_EVENTS) == DF_ON ;
NABoolean possibleMultipleCalls = generator->getRightSideOfFlow() ;
short scratchTresholdPct =
(short) CmpCommon::getDefaultLong(SCRATCH_FREESPACE_THRESHOLD_PERCENT);
// determione the memory usage (amount of memory as percentage from total
// physical memory used to initialize data structures)
unsigned short memUsagePercent =
(unsigned short) getDefault(BMO_MEMORY_USAGE_PERCENT);
short memPressurePct = (short)getDefault(GEN_MEM_PRESSURE_THRESHOLD);
historyRecLen = ROUND8(historyRecLen);
Lng32 maxNumberOfOLAPBuffers;
Lng32 maxRowsInOLAPBuffer;
Lng32 minNumberOfOLAPBuffers;
Lng32 numberOfWinOLAPBuffers;
Lng32 olapBufferSize;
computeHistoryParams(historyRecLen,
maxRowsInOLAPBuffer,
minNumberOfOLAPBuffers,
numberOfWinOLAPBuffers,
maxNumberOfOLAPBuffers,
olapBufferSize);
ComTdbSequence *sequenceTdb
= new(space) ComTdbSequence(readSeqExpr,
returnExpr,
postPred,
cancelExpression,
getMinFollowingRows(),
#pragma nowarn(1506) // warning elimination
historyRecLen,
historyAtpIndex,
childTdb,
givenCriDesc,
returnCriDesc,
(queue_index)getDefault(GEN_SEQFUNC_SIZE_DOWN),
(queue_index)getDefault(GEN_SEQFUNC_SIZE_UP),
getDefault(GEN_SEQFUNC_NUM_BUFFERS),
getDefault(GEN_SEQFUNC_BUFFER_SIZE),
olapBufferSize,
maxNumberOfOLAPBuffers,
numHistoryRows(),
getUnboundedFollowing(),
logDiagnostics,
possibleMultipleCalls,
scratchTresholdPct,
memUsagePercent,
memPressurePct,
maxRowsInOLAPBuffer,
minNumberOfOLAPBuffers,
numberOfWinOLAPBuffers,
noOverflow,
checkPartChangeExpr);
#pragma warn(1506) // warning elimination
generator->initTdbFields(sequenceTdb);
// update the estimated value of HistoryRowLength with actual value
//setEstHistoryRowLength(historyIds.getRowLength());
double sequenceMemEst = getEstimatedRunTimeMemoryUsage(sequenceTdb);
generator->addToTotalEstimatedMemory(sequenceMemEst);
if(!generator->explainDisabled()) {
Lng32 seqMemEstInKBPerCPU = (Lng32)(sequenceMemEst / 1024) ;
seqMemEstInKBPerCPU = seqMemEstInKBPerCPU/
(MAXOF(generator->compilerStatsInfo().dop(),1));
generator->setOperEstimatedMemory(seqMemEstInKBPerCPU);
generator->
setExplainTuple(addExplainInfo(sequenceTdb,
childExplainTuple,
0,
generator));
generator->setOperEstimatedMemory(0);
}
sequenceTdb->setScratchIOVectorSize((Int16)getDefault(SCRATCH_IO_VECTOR_SIZE_HASH));
sequenceTdb->setOverflowMode(generator->getOverflowMode());
sequenceTdb->setBmoMinMemBeforePressureCheck((Int16)getDefault(EXE_BMO_MIN_SIZE_BEFORE_PRESSURE_CHECK_IN_MB));
if(generator->getOverflowMode() == ComTdb::OFM_SSD )
sequenceTdb->setBMOMaxMemThresholdMB((UInt16)(ActiveSchemaDB()->
getDefaults()).
getAsLong(SSD_BMO_MAX_MEM_THRESHOLD_IN_MB));
else
sequenceTdb->setBMOMaxMemThresholdMB((UInt16)(ActiveSchemaDB()->
getDefaults()).
getAsLong(EXE_MEMORY_AVAILABLE_IN_MB));
// The CQD EXE_MEM_LIMIT_PER_BMO_IN_MB has precedence over the mem quota sys
NADefaults &defs = ActiveSchemaDB()->getDefaults();
UInt16 mmu = (UInt16)(defs.getAsDouble(EXE_MEM_LIMIT_PER_BMO_IN_MB));
UInt16 numBMOsInFrag = (UInt16)generator->getFragmentDir()->getNumBMOs();
if (mmu != 0)
sequenceTdb->setMemoryQuotaMB(mmu);
else {
// Apply quota system if either one the following two is true:
// 1. the memory limit feature is turned off and more than one BMOs
// 2. the memory limit feature is turned on
NABoolean mlimitPerCPU = defs.getAsDouble(EXE_MEMORY_LIMIT_PER_CPU) > 0;
if ( mlimitPerCPU || numBMOsInFrag > 1 ) {
double memQuota =
computeMemoryQuota(generator->getEspLevel() == 0,
mlimitPerCPU,
generator->getBMOsMemoryLimitPerCPU().value(),
generator->getTotalNumBMOsPerCPU(),
generator->getTotalBMOsMemoryPerCPU().value(),
numBMOsInFrag,
generator->getFragmentDir()->getBMOsMemoryUsage()
);
sequenceTdb->setMemoryQuotaMB( UInt16(memQuota) );
}
}
generator->setCriDesc(givenCriDesc, Generator::DOWN);
generator->setCriDesc(returnCriDesc, Generator::UP);
generator->setGenObj(this, sequenceTdb);
return 0;
}
// AppliedStatMan::setupASMCacheForJBB method will be called from
// Query::Analyze after connectivity analysis has been done and
// empty logical properties have been set.
void AppliedStatMan::setupASMCacheForJBB(JBB & jbb)
{
EstLogPropSharedPtr myEstLogProp;
// get all JBBCs of JBB
const CANodeIdSet jbbcNodeIdSet = jbb.getMainJBBSubset().getJBBCs();
CANodeId jbbcId;
// for all jbbcs
for (jbbcId = jbbcNodeIdSet.init();
jbbcNodeIdSet.next(jbbcId);
jbbcNodeIdSet.advance(jbbcId))
{
if (NodeAnalysis * jbbcNode = jbbcId.getNodeAnalysis())
{
// Evaluate local predicates only if it is a table.
RelExpr * jbbcExpr = jbbcNode->getOriginalExpr();
if ((jbbcNode->getTableAnalysis() != NULL) &&
(jbbcExpr->getOperatorType() == REL_SCAN))
{
// get the original expression of the jbbc
Scan * scanExpr = (Scan *) jbbcExpr;
ValueIdSet localPreds = scanExpr->getSelectionPredicates();
// if local predicates have already been computed, then skip
if ((localPreds.entries() > 0) || !(lookup(jbbcId)))
{
// check to see this GA has already been associated with
// a logExpr for synthesis. If not, then synthesize
// log. expression, and then apply local predicates to it
if (NOT scanExpr->getGroupAttr()->existsLogExprForSynthesis())
scanExpr->synthLogProp();
myEstLogProp = getStatsForCANodeId(jbbcId);
}
}
}
}
// Now do a second traversal of the JBB looking for join reducers
for (jbbcId = jbbcNodeIdSet.init();
jbbcNodeIdSet.next(jbbcId);
jbbcNodeIdSet.advance(jbbcId))
{
// now look for all two way joins for this child
if (jbbcId.getNodeAnalysis())
{
// get all JBBCs connected to this JBBC, and do a two-way
// join with all of them
CANodeIdSet connectedNodes = jbbcId.getNodeAnalysis()->\
getJBBC()->getJoinedJBBCs();
for (CANodeId connectedTable = connectedNodes.init();
connectedNodes.next(connectedTable);
connectedNodes.advance(connectedTable))
{
if (connectedTable.getNodeAnalysis())
{
// ASM does not concern itself with the order of the tables,
// hence it is possible that the join has already been computed
CANodeIdSet tableSet = jbbcId;
tableSet.insert(connectedTable);
if ((myEstLogProp = getCachedStatistics(&tableSet)) == NULL)
{
CANodeIdSet setForjbbcId(jbbcId);
CANodeIdSet setForConnectedTable(connectedTable);
myEstLogProp = joinJBBChildren(setForjbbcId, setForConnectedTable);
}
}
}
}
}
} // AppliedStatMan::setupASMCacheForJBB
