public ExprNodeGenericFuncEvaluator(ExprNodeGenericFuncDesc expr) {
this.expr = expr;
children = new ExprNodeEvaluator[expr.getChildExprs().size()];
isEager = false;
for (int i = 0; i < children.length; i++) {
ExprNodeDesc child = expr.getChildExprs().get(i);
ExprNodeEvaluator nodeEvaluator = ExprNodeEvaluatorFactory.get(child);
children[i] = nodeEvaluator;
// If we have eager evaluators anywhere below us, then we are eager too.
if (nodeEvaluator instanceof ExprNodeGenericFuncEvaluator) {
if (((ExprNodeGenericFuncEvaluator) nodeEvaluator).isEager) {
isEager = true;
}
// Base case: we are eager if a child is stateful
GenericUDF childUDF = ((ExprNodeGenericFuncDesc) child).getGenericUDF();
if (FunctionRegistry.isStateful(childUDF)) {
isEager = true;
}
}
}
deferredChildren = new GenericUDF.DeferredObject[expr.getChildExprs().size()];
for (int i = 0; i < deferredChildren.length; i++) {
if (isEager) {
deferredChildren[i] = new EagerExprObject(children[i]);
} else {
deferredChildren[i] = new DeferredExprObject(children[i]);
}
}
}
// Check if ExprNodeColumnDesc is wrapped in expr.
// If so, peel off. Otherwise return itself.
private ExprNodeDesc getColumnExpr(ExprNodeDesc expr) {
if (expr instanceof ExprNodeColumnDesc) {
return expr;
}
ExprNodeGenericFuncDesc funcDesc = null;
if (expr instanceof ExprNodeGenericFuncDesc) {
funcDesc = (ExprNodeGenericFuncDesc) expr;
}
if (null == funcDesc) {
return expr;
}
GenericUDF udf = funcDesc.getGenericUDF();
// check if its a simple cast expression.
if ((udf instanceof GenericUDFBridge
|| udf instanceof GenericUDFToBinary
|| udf instanceof GenericUDFToChar
|| udf instanceof GenericUDFToVarchar
|| udf instanceof GenericUDFToDecimal
|| udf instanceof GenericUDFToDate
|| udf instanceof GenericUDFToUnixTimeStamp
|| udf instanceof GenericUDFToUtcTimestamp)
&& funcDesc.getChildren().size() == 1
&& funcDesc.getChildren().get(0) instanceof ExprNodeColumnDesc) {
return expr.getChildren().get(0);
}
return expr;
}
private RexNode handleExplicitCast(ExprNodeGenericFuncDesc func, List<RexNode> childRexNodeLst)
throws CalciteSemanticException {
RexNode castExpr = null;
if (childRexNodeLst != null && childRexNodeLst.size() == 1) {
GenericUDF udf = func.getGenericUDF();
if ((udf instanceof GenericUDFToChar)
|| (udf instanceof GenericUDFToVarchar)
|| (udf instanceof GenericUDFToDecimal)
|| (udf instanceof GenericUDFToDate)
|| (udf instanceof GenericUDFToBinary)
|| castExprUsingUDFBridge(udf)) {
castExpr =
cluster
.getRexBuilder()
.makeAbstractCast(
TypeConverter.convert(func.getTypeInfo(), cluster.getTypeFactory()),
childRexNodeLst.get(0));
}
}
return castExpr;
}
@Override
public ObjectInspector initialize(ObjectInspector rowInspector) throws HiveException {
// Initialize all children first
ObjectInspector[] childrenOIs = new ObjectInspector[children.length];
for (int i = 0; i < children.length; i++) {
childrenOIs[i] = children[i].initialize(rowInspector);
}
genericUDF = expr.getGenericUDF();
if (isEager
&& ((genericUDF instanceof GenericUDFCase) || (genericUDF instanceof GenericUDFWhen))) {
throw new HiveException("Stateful expressions cannot be used inside of CASE");
}
this.outputOI = genericUDF.initializeAndFoldConstants(childrenOIs);
return this.outputOI;
}
/**
* If the genericUDF is a base comparison, it returns an integer based on the result of comparing
* the two sides of the UDF, like the compareTo method in Comparable.
*
* <p>If the genericUDF is not a base comparison, or there is an error executing the comparison,
* it returns null.
*
* @param row
* @return
* @throws HiveException
*/
public Integer compare(Object row) throws HiveException {
if (!expr.isSortedExpr() || !(genericUDF instanceof GenericUDFBaseCompare)) {
for (ExprNodeEvaluator evaluator : children) {
if (evaluator instanceof ExprNodeGenericFuncEvaluator) {
Integer comparison = ((ExprNodeGenericFuncEvaluator) evaluator).compare(row);
if (comparison != null) {
return comparison;
}
}
}
return null;
}
rowObject = row;
if (isEager) {
for (int i = 0; i < deferredChildren.length; i++) {
((EagerExprObject) deferredChildren[i]).evaluate();
}
}
return ((GenericUDFBaseCompare) genericUDF).compare(deferredChildren);
}
private ExprNodeDesc analyzeExpr(
ExprNodeGenericFuncDesc expr,
List<IndexSearchCondition> searchConditions,
Object... nodeOutputs) {
if (FunctionRegistry.isOpAnd(expr)) {
assert (nodeOutputs.length == 2);
ExprNodeDesc residual1 = (ExprNodeDesc) nodeOutputs[0];
ExprNodeDesc residual2 = (ExprNodeDesc) nodeOutputs[1];
if (residual1 == null) {
return residual2;
}
if (residual2 == null) {
return residual1;
}
List<ExprNodeDesc> residuals = new ArrayList<ExprNodeDesc>();
residuals.add(residual1);
residuals.add(residual2);
return new ExprNodeGenericFuncDesc(
TypeInfoFactory.booleanTypeInfo, FunctionRegistry.getGenericUDFForAnd(), residuals);
}
GenericUDF genericUDF = expr.getGenericUDF();
if (!(genericUDF instanceof GenericUDFBaseCompare)) {
return expr;
}
ExprNodeDesc expr1 = (ExprNodeDesc) nodeOutputs[0];
ExprNodeDesc expr2 = (ExprNodeDesc) nodeOutputs[1];
// We may need to peel off the GenericUDFBridge that is added by CBO or user
if (expr1.getTypeInfo().equals(expr2.getTypeInfo())) {
expr1 = getColumnExpr(expr1);
expr2 = getColumnExpr(expr2);
}
ExprNodeDesc[] extracted = ExprNodeDescUtils.extractComparePair(expr1, expr2);
if (extracted == null || (extracted.length > 2 && !acceptsFields)) {
return expr;
}
ExprNodeColumnDesc columnDesc;
ExprNodeConstantDesc constantDesc;
if (extracted[0] instanceof ExprNodeConstantDesc) {
genericUDF = genericUDF.flip();
columnDesc = (ExprNodeColumnDesc) extracted[1];
constantDesc = (ExprNodeConstantDesc) extracted[0];
} else {
columnDesc = (ExprNodeColumnDesc) extracted[0];
constantDesc = (ExprNodeConstantDesc) extracted[1];
}
String udfName = genericUDF.getUdfName();
if (!udfNames.contains(genericUDF.getUdfName())) {
return expr;
}
if (!allowedColumnNames.contains(columnDesc.getColumn())) {
return expr;
}
String[] fields = null;
if (extracted.length > 2) {
ExprNodeFieldDesc fieldDesc = (ExprNodeFieldDesc) extracted[2];
if (!isValidField(fieldDesc)) {
return expr;
}
fields = ExprNodeDescUtils.extractFields(fieldDesc);
}
// We also need to update the expr so that the index query can be generated.
// Note that, hive does not support UDFToDouble etc in the query text.
List<ExprNodeDesc> list = new ArrayList<ExprNodeDesc>();
list.add(expr1);
list.add(expr2);
expr = new ExprNodeGenericFuncDesc(expr.getTypeInfo(), expr.getGenericUDF(), list);
searchConditions.add(new IndexSearchCondition(columnDesc, udfName, constantDesc, expr, fields));
// we converted the expression to a search condition, so
// remove it from the residual predicate
return fields == null ? null : expr;
}
/**
* Construct the filter expression from the skewed keys and skewed values. If the skewed join
* keys are (k1), and (k1,k3) with the skewed values (1,2) and ((2,3),(4,5)) respectively, the
* filter expression would be: (k1=1) or (k1=2) or ((k1=2) and (k3=3)) or ((k1=4) and (k3=5)).
*/
private ExprNodeDesc constructFilterExpr(
Map<List<ExprNodeDesc>, List<List<String>>> skewedValuesMap, boolean skewed) {
ExprNodeDesc finalExprNodeDesc = null;
try {
for (Map.Entry<List<ExprNodeDesc>, List<List<String>>> mapEntry :
skewedValuesMap.entrySet()) {
List<ExprNodeDesc> keyCols = mapEntry.getKey();
List<List<String>> skewedValuesList = mapEntry.getValue();
for (List<String> skewedValues : skewedValuesList) {
int keyPos = 0;
ExprNodeDesc currExprNodeDesc = null;
// Make the following condition: all the values match for all the columns
for (String skewedValue : skewedValues) {
List<ExprNodeDesc> children = new ArrayList<ExprNodeDesc>();
// We have ensured that the keys are columns
ExprNodeColumnDesc keyCol = (ExprNodeColumnDesc) keyCols.get(keyPos).clone();
keyPos++;
children.add(keyCol);
// Convert the constants available as strings to the corresponding objects
children.add(createConstDesc(skewedValue, keyCol));
ExprNodeGenericFuncDesc expr = null;
// Create the equality condition
expr = ExprNodeGenericFuncDesc.newInstance(new GenericUDFOPEqual(), children);
if (currExprNodeDesc == null) {
currExprNodeDesc = expr;
} else {
// If there are previous nodes, then AND the current node with the previous one
List<ExprNodeDesc> childrenAND = new ArrayList<ExprNodeDesc>();
childrenAND.add(currExprNodeDesc);
childrenAND.add(expr);
currExprNodeDesc =
ExprNodeGenericFuncDesc.newInstance(new GenericUDFOPAnd(), childrenAND);
}
}
// If there are more than one skewed values,
// then OR the current node with the previous one
if (finalExprNodeDesc == null) {
finalExprNodeDesc = currExprNodeDesc;
} else {
List<ExprNodeDesc> childrenOR = new ArrayList<ExprNodeDesc>();
childrenOR.add(finalExprNodeDesc);
childrenOR.add(currExprNodeDesc);
finalExprNodeDesc =
ExprNodeGenericFuncDesc.newInstance(new GenericUDFOPOr(), childrenOR);
}
}
}
// Add a NOT operator in the beginning (this is for the cloned operator because we
// want the values which are not skewed
if (skewed == false) {
List<ExprNodeDesc> childrenNOT = new ArrayList<ExprNodeDesc>();
childrenNOT.add(finalExprNodeDesc);
finalExprNodeDesc =
ExprNodeGenericFuncDesc.newInstance(new GenericUDFOPNot(), childrenNOT);
}
} catch (UDFArgumentException e) {
// Ignore the exception because we are not comparing Long vs. String here.
// There should never be an exception
assert false;
}
return finalExprNodeDesc;
}
private RexNode convert(final ExprNodeGenericFuncDesc func) throws SemanticException {
ExprNodeDesc tmpExprNode;
RexNode tmpRN;
List<RexNode> childRexNodeLst = new LinkedList<RexNode>();
Builder<RelDataType> argTypeBldr = ImmutableList.<RelDataType>builder();
// TODO: 1) Expand to other functions as needed 2) What about types other than primitive.
TypeInfo tgtDT = null;
GenericUDF tgtUdf = func.getGenericUDF();
boolean isNumeric =
(tgtUdf instanceof GenericUDFBaseBinary
&& func.getTypeInfo().getCategory() == Category.PRIMITIVE
&& (PrimitiveGrouping.NUMERIC_GROUP
== PrimitiveObjectInspectorUtils.getPrimitiveGrouping(
((PrimitiveTypeInfo) func.getTypeInfo()).getPrimitiveCategory())));
boolean isCompare = !isNumeric && tgtUdf instanceof GenericUDFBaseCompare;
if (isNumeric) {
tgtDT = func.getTypeInfo();
assert func.getChildren().size() == 2;
// TODO: checking 2 children is useless, compare already does that.
} else if (isCompare && (func.getChildren().size() == 2)) {
tgtDT =
FunctionRegistry.getCommonClassForComparison(
func.getChildren().get(0).getTypeInfo(), func.getChildren().get(1).getTypeInfo());
}
for (ExprNodeDesc childExpr : func.getChildren()) {
tmpExprNode = childExpr;
if (tgtDT != null
&& TypeInfoUtils.isConversionRequiredForComparison(tgtDT, childExpr.getTypeInfo())) {
if (isCompare) {
// For compare, we will convert requisite children
tmpExprNode = ParseUtils.createConversionCast(childExpr, (PrimitiveTypeInfo) tgtDT);
} else if (isNumeric) {
// For numeric, we'll do minimum necessary cast - if we cast to the type
// of expression, bad things will happen.
PrimitiveTypeInfo minArgType = ExprNodeDescUtils.deriveMinArgumentCast(childExpr, tgtDT);
tmpExprNode = ParseUtils.createConversionCast(childExpr, minArgType);
} else {
throw new AssertionError("Unexpected " + tgtDT + " - not a numeric op or compare");
}
}
argTypeBldr.add(TypeConverter.convert(tmpExprNode.getTypeInfo(), cluster.getTypeFactory()));
tmpRN = convert(tmpExprNode);
childRexNodeLst.add(tmpRN);
}
// See if this is an explicit cast.
RexNode expr = null;
RelDataType retType = null;
expr = handleExplicitCast(func, childRexNodeLst);
if (expr == null) {
// This is not a cast; process the function.
retType = TypeConverter.convert(func.getTypeInfo(), cluster.getTypeFactory());
SqlOperator calciteOp =
SqlFunctionConverter.getCalciteOperator(
func.getFuncText(), func.getGenericUDF(), argTypeBldr.build(), retType);
expr = cluster.getRexBuilder().makeCall(calciteOp, childRexNodeLst);
} else {
retType = expr.getType();
}
// TODO: Cast Function in Calcite have a bug where it infer type on cast throws
// an exception
if (flattenExpr
&& (expr instanceof RexCall)
&& !(((RexCall) expr).getOperator() instanceof SqlCastFunction)) {
RexCall call = (RexCall) expr;
expr =
cluster
.getRexBuilder()
.makeCall(
retType,
call.getOperator(),
RexUtil.flatten(call.getOperands(), call.getOperator()));
}
return expr;
}