Merge remote-tracking branch 'apache/master' into SPARK-23951

# Conflicts:
#	sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateUnsafeProjection.scala

Author: hvanhovell

mllib/src/main/scala/org/apache/spark/ml/tree/impl/NodeIdCache.scala

@@ -95,7 +95,7 @@ private[spark] class NodeIdCache(
       splits: Array[Array[Split]]): Unit = {
     if (prevNodeIdsForInstances != null) {
       // Unpersist the previous one if one exists.
-      prevNodeIdsForInstances.unpersist()
+      prevNodeIdsForInstances.unpersist(false)
     }
 
     prevNodeIdsForInstances = nodeIdsForInstances
@@ -166,9 +166,13 @@ private[spark] class NodeIdCache(
         }
       }
     }
+    if (nodeIdsForInstances != null) {
+      // Unpersist current one if one exists.
+      nodeIdsForInstances.unpersist(false)
+    }
     if (prevNodeIdsForInstances != null) {
       // Unpersist the previous one if one exists.
-      prevNodeIdsForInstances.unpersist()
+      prevNodeIdsForInstances.unpersist(false)
     }
   }
 }

sql/catalyst/src/main/java/org/apache/spark/sql/catalyst/expressions/codegen/UnsafeWriter.java

@@ -16,6 +16,9 @@
  */
 package org.apache.spark.sql.catalyst.expressions.codegen;
 
+import org.apache.spark.sql.catalyst.expressions.UnsafeArrayData;
+import org.apache.spark.sql.catalyst.expressions.UnsafeMapData;
+import org.apache.spark.sql.catalyst.expressions.UnsafeRow;
 import org.apache.spark.sql.types.Decimal;
 import org.apache.spark.unsafe.Platform;
 import org.apache.spark.unsafe.array.ByteArrayMethods;
@@ -103,42 +106,27 @@ protected final void zeroOutPaddingBytes(int numBytes) {
   public abstract void write(int ordinal, Decimal input, int precision, int scale);
 
   public final void write(int ordinal, UTF8String input) {
-    final int numBytes = input.numBytes();
-    final int roundedSize = ByteArrayMethods.roundNumberOfBytesToNearestWord(numBytes);
-
-    // grow the global buffer before writing data.
-    grow(roundedSize);
-
-    zeroOutPaddingBytes(numBytes);
-
-    // Write the bytes to the variable length portion.
-    input.writeToMemory(getBuffer(), cursor());
-
-    setOffsetAndSize(ordinal, numBytes);
-
-    // move the cursor forward.
-    increaseCursor(roundedSize);
+    writeUnalignedBytes(ordinal, input.getBaseObject(), input.getBaseOffset(), input.numBytes());
   }
 
   public final void write(int ordinal, byte[] input) {
     write(ordinal, input, 0, input.length);
   }
 
   public final void write(int ordinal, byte[] input, int offset, int numBytes) {
-    final int roundedSize = ByteArrayMethods.roundNumberOfBytesToNearestWord(input.length);
+    writeUnalignedBytes(ordinal, input, Platform.BYTE_ARRAY_OFFSET + offset, numBytes);
+  }
 
-    // grow the global buffer before writing data.
+  private void writeUnalignedBytes(
+      int ordinal,
+      Object baseObject,
+      long baseOffset,
+      int numBytes) {
+    final int roundedSize = ByteArrayMethods.roundNumberOfBytesToNearestWord(numBytes);
     grow(roundedSize);
-
     zeroOutPaddingBytes(numBytes);
-
-    // Write the bytes to the variable length portion.
-    Platform.copyMemory(
-      input, Platform.BYTE_ARRAY_OFFSET + offset, getBuffer(), cursor(), numBytes);
-
+    Platform.copyMemory(baseObject, baseOffset, getBuffer(), cursor(), numBytes);
     setOffsetAndSize(ordinal, numBytes);
-
-    // move the cursor forward.
     increaseCursor(roundedSize);
   }
 
@@ -156,6 +144,40 @@ public final void write(int ordinal, CalendarInterval input) {
     increaseCursor(16);
   }
 
+  public final void write(int ordinal, UnsafeRow row) {
+    writeAlignedBytes(ordinal, row.getBaseObject(), row.getBaseOffset(), row.getSizeInBytes());
+  }
+
+  public final void write(int ordinal, UnsafeMapData map) {
+    writeAlignedBytes(ordinal, map.getBaseObject(), map.getBaseOffset(), map.getSizeInBytes());
+  }
+
+  public final void write(UnsafeArrayData array) {
+    // Unsafe arrays both can be written as a regular array field or as part of a map. This makes
+    // updating the offset and size dependent on the code path, this is why we currently do not
+    // provide an method for writing unsafe arrays that also updates the size and offset.
+    int numBytes = array.getSizeInBytes();
+    grow(numBytes);
+    Platform.copyMemory(
+            array.getBaseObject(),
+            array.getBaseOffset(),
+            getBuffer(),
+            cursor(),
+            numBytes);
+    increaseCursor(numBytes);
+  }
+
+  private void writeAlignedBytes(
+      int ordinal,
+      Object baseObject,
+      long baseOffset,
+      int numBytes) {
+    grow(numBytes);
+    Platform.copyMemory(baseObject, baseOffset, getBuffer(), cursor(), numBytes);
+    setOffsetAndSize(ordinal, numBytes);
+    increaseCursor(numBytes);
+  }
+
   protected final void writeBoolean(long offset, boolean value) {
     Platform.putBoolean(getBuffer(), offset, value);
   }

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/InterpretedUnsafeProjection.scala

@@ -173,21 +173,17 @@ object InterpretedUnsafeProjection extends UnsafeProjectionCreator {
         val rowWriter = new UnsafeRowWriter(writer, numFields)
         val structWriter = generateStructWriter(rowWriter, fields)
         (v, i) => {
-          val previousCursor = writer.cursor()
           v.getStruct(i, fields.length) match {
             case row: UnsafeRow =>
-              writeUnsafeData(
-                rowWriter,
-                row.getBaseObject,
-                row.getBaseOffset,
-                row.getSizeInBytes)
+              writer.write(i, row)
             case row =>
+              val previousCursor = writer.cursor()
               // Nested struct. We don't know where this will start because a row can be
               // variable length, so we need to update the offsets and zero out the bit mask.
               rowWriter.resetRowWriter()
               structWriter.apply(row)
+              writer.setOffsetAndSizeFromPreviousCursor(i, previousCursor)
           }
-          writer.setOffsetAndSizeFromPreviousCursor(i, previousCursor)
         }
 
       case ArrayType(elementType, containsNull) =>
@@ -214,15 +210,12 @@ object InterpretedUnsafeProjection extends UnsafeProjectionCreator {
           valueType,
           valueContainsNull)
         (v, i) => {
-          val previousCursor = writer.cursor()
           v.getMap(i) match {
             case map: UnsafeMapData =>
-              writeUnsafeData(
-                valueArrayWriter,
-                map.getBaseObject,
-                map.getBaseOffset,
-                map.getSizeInBytes)
+              writer.write(i, map)
             case map =>
+              val previousCursor = writer.cursor()
+
               // preserve 8 bytes to write the key array numBytes later.
               valueArrayWriter.grow(8)
               valueArrayWriter.increaseCursor(8)
@@ -237,8 +230,8 @@ object InterpretedUnsafeProjection extends UnsafeProjectionCreator {
 
               // Write the values.
               writeArray(valueArrayWriter, valueWriter, map.valueArray())
+              writer.setOffsetAndSizeFromPreviousCursor(i, previousCursor)
           }
-          writer.setOffsetAndSizeFromPreviousCursor(i, previousCursor)
         }
 
       case udt: UserDefinedType[_] =>
@@ -318,11 +311,7 @@ object InterpretedUnsafeProjection extends UnsafeProjectionCreator {
       elementWriter: (SpecializedGetters, Int) => Unit,
       array: ArrayData): Unit = array match {
     case unsafe: UnsafeArrayData =>
-      writeUnsafeData(
-        arrayWriter,
-        unsafe.getBaseObject,
-        unsafe.getBaseOffset,
-        unsafe.getSizeInBytes)
+      arrayWriter.write(unsafe)
     case _ =>
       val numElements = array.numElements()
       arrayWriter.initialize(numElements)
@@ -332,23 +321,4 @@ object InterpretedUnsafeProjection extends UnsafeProjectionCreator {
         i += 1
       }
   }
-
-  /**
-   * Write an opaque block of data to the buffer. This is used to copy
-   * [[UnsafeRow]], [[UnsafeArrayData]] and [[UnsafeMapData]] objects.
-   */
-  private def writeUnsafeData(
-      writer: UnsafeWriter,
-      baseObject: AnyRef,
-      baseOffset: Long,
-      sizeInBytes: Int) : Unit = {
-    writer.grow(sizeInBytes)
-    Platform.copyMemory(
-      baseObject,
-      baseOffset,
-      writer.getBuffer,
-      writer.cursor,
-      sizeInBytes)
-    writer.increaseCursor(sizeInBytes)
-  }
 }