Github user kiszk commented on a diff in the pull request:
https://github.com/apache/spark/pull/21021#discussion_r186060660
--- Diff:
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/collectionOperations.scala
---
@@ -191,28 +161,191 @@ case class SortArray(base: Expression,
ascendingOrder: Expression)
if (o1 == null && o2 == null) {
0
} else if (o1 == null) {
- 1
+ -nullOrder
} else if (o2 == null) {
- -1
+ nullOrder
} else {
-ordering.compare(o1, o2)
}
}
}
}
- override def nullSafeEval(array: Any, ascending: Any): Any = {
- val elementType = base.dataType.asInstanceOf[ArrayType].elementType
+ def elementType: DataType =
arrayExpression.dataType.asInstanceOf[ArrayType].elementType
+
+ def sortEval(array: Any, ascending: Boolean): Any = {
val data = array.asInstanceOf[ArrayData].toArray[AnyRef](elementType)
if (elementType != NullType) {
- java.util.Arrays.sort(data, if (ascending.asInstanceOf[Boolean]) lt
else gt)
+ java.util.Arrays.sort(data, if (ascending) lt else gt)
}
new GenericArrayData(data.asInstanceOf[Array[Any]])
}
+ def sortCodegen(ctx: CodegenContext, ev: ExprCode, base: String, order:
String): String = {
+ val arrayData = classOf[ArrayData].getName
+ val genericArrayData = classOf[GenericArrayData].getName
+ val array = ctx.freshName("array")
+ val c = ctx.freshName("c")
+ if (elementType == NullType) {
+ s"${ev.value} = $base.copy();"
+ } else {
+ val elementTypeTerm = ctx.addReferenceObj("elementTypeTerm",
elementType)
+ val sortOrder = ctx.freshName("sortOrder")
+ val o1 = ctx.freshName("o1")
+ val o2 = ctx.freshName("o2")
+ val jt = CodeGenerator.javaType(elementType)
+ val comp = if (CodeGenerator.isPrimitiveType(elementType)) {
+ val bt = CodeGenerator.boxedType(elementType)
+ val v1 = ctx.freshName("v1")
+ val v2 = ctx.freshName("v2")
+ s"""
+ |$jt $v1 = (($bt) $o1).${jt}Value();
--- End diff --
IIUC, this is because `compare()` in `java.util.Arrays.sort` accepts two
`Object` arguments. If we could use existing type specialized sort algorithm
(i.e. compare(int, int), compare(float, float), and so one), we are happy to
avoid boxing.
Do you have any suggestion for such an existing sorting implementation?
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
