[SPARK-18016][SQL][CATALYST][BRANCH-2.2] Code Generation: Constant Pool Limit - Class Splitting

## What changes were proposed in this pull request?

This is a backport patch for Spark 2.2.x of the class splitting feature over excess generated code as was merged in apache#18075.

## How was this patch tested?

The same test provided in apache#18075 is included in this patch.

Author: ALeksander Eskilson <[email protected]>

Closes apache#18377 from bdrillard/class_splitting_2.2.

Author: ALeksander Eskilson

sql/catalyst/pom.xml

@@ -131,6 +131,13 @@
           </execution>
         </executions>
       </plugin>
+      <plugin>
+        <groupId>org.scalatest</groupId>
+        <artifactId>scalatest-maven-plugin</artifactId>
+        <configuration>
+          <argLine>-Xmx4g -Xss4096k -XX:MaxPermSize=${MaxPermGen} -XX:ReservedCodeCacheSize=512m</argLine>
+        </configuration>
+      </plugin>
       <plugin>
         <groupId>org.antlr</groupId>
         <artifactId>antlr4-maven-plugin</artifactId>

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

@@ -988,7 +988,7 @@ case class ScalaUDF(
     val converterTerm = ctx.freshName("converter")
     val expressionIdx = ctx.references.size - 1
     ctx.addMutableState(converterClassName, converterTerm,
-      s"this.$converterTerm = ($converterClassName)$typeConvertersClassName" +
+      s"$converterTerm = ($converterClassName)$typeConvertersClassName" +
         s".createToScalaConverter(((${expressionClassName})((($scalaUDFClassName)" +
           s"references[$expressionIdx]).getChildren().apply($index))).dataType());")
     converterTerm
@@ -1005,7 +1005,7 @@ case class ScalaUDF(
     // Generate codes used to convert the returned value of user-defined functions to Catalyst type
     val catalystConverterTerm = ctx.freshName("catalystConverter")
     ctx.addMutableState(converterClassName, catalystConverterTerm,
-      s"this.$catalystConverterTerm = ($converterClassName)$typeConvertersClassName" +
+      s"$catalystConverterTerm = ($converterClassName)$typeConvertersClassName" +
         s".createToCatalystConverter($scalaUDF.dataType());")
 
     val resultTerm = ctx.freshName("result")
@@ -1019,7 +1019,7 @@ case class ScalaUDF(
 
     val funcTerm = ctx.freshName("udf")
     ctx.addMutableState(funcClassName, funcTerm,
-      s"this.$funcTerm = ($funcClassName)$scalaUDF.userDefinedFunc();")
+      s"$funcTerm = ($funcClassName)$scalaUDF.userDefinedFunc();")
 
     // codegen for children expressions
     val evals = children.map(_.genCode(ctx))

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/CodeGenerator.scala

@@ -113,7 +113,7 @@ class CodegenContext {
     val idx = references.length
     references += obj
     val clsName = Option(className).getOrElse(obj.getClass.getName)
-    addMutableState(clsName, term, s"this.$term = ($clsName) references[$idx];")
+    addMutableState(clsName, term, s"$term = ($clsName) references[$idx];")
     term
   }
 
@@ -202,16 +202,6 @@ class CodegenContext {
     partitionInitializationStatements.mkString("\n")
   }
 
-  /**
-   * Holding all the functions those will be added into generated class.
-   */
-  val addedFunctions: mutable.Map[String, String] =
-    mutable.Map.empty[String, String]
-
-  def addNewFunction(funcName: String, funcCode: String): Unit = {
-    addedFunctions += ((funcName, funcCode))
-  }
-
   /**
    * Holds expressions that are equivalent. Used to perform subexpression elimination
    * during codegen.
@@ -233,10 +223,118 @@ class CodegenContext {
   // The collection of sub-expression result resetting methods that need to be called on each row.
   val subexprFunctions = mutable.ArrayBuffer.empty[String]
 
-  def declareAddedFunctions(): String = {
-    addedFunctions.map { case (funcName, funcCode) => funcCode }.mkString("\n")
+  private val outerClassName = "OuterClass"
+
+  /**
+   * Holds the class and instance names to be generated, where `OuterClass` is a placeholder
+   * standing for whichever class is generated as the outermost class and which will contain any
+   * nested sub-classes. All other classes and instance names in this list will represent private,
+   * nested sub-classes.
+   */
+  private val classes: mutable.ListBuffer[(String, String)] =
+    mutable.ListBuffer[(String, String)](outerClassName -> null)
+
+  // A map holding the current size in bytes of each class to be generated.
+  private val classSize: mutable.Map[String, Int] =
+    mutable.Map[String, Int](outerClassName -> 0)
+
+  // Nested maps holding function names and their code belonging to each class.
+  private val classFunctions: mutable.Map[String, mutable.Map[String, String]] =
+    mutable.Map(outerClassName -> mutable.Map.empty[String, String])
+
+  // Returns the size of the most recently added class.
+  private def currClassSize(): Int = classSize(classes.head._1)
+
+  // Returns the class name and instance name for the most recently added class.
+  private def currClass(): (String, String) = classes.head
+
+  // Adds a new class. Requires the class' name, and its instance name.
+  private def addClass(className: String, classInstance: String): Unit = {
+    classes.prepend(className -> classInstance)
+    classSize += className -> 0
+    classFunctions += className -> mutable.Map.empty[String, String]
   }
 
+  /**
+   * Adds a function to the generated class. If the code for the `OuterClass` grows too large, the
+   * function will be inlined into a new private, nested class, and a instance-qualified name for
+   * the function will be returned. Otherwise, the function will be inined to the `OuterClass` the
+   * simple `funcName` will be returned.
+   *
+   * @param funcName the class-unqualified name of the function
+   * @param funcCode the body of the function
+   * @param inlineToOuterClass whether the given code must be inlined to the `OuterClass`. This
+   *                           can be necessary when a function is declared outside of the context
+   *                           it is eventually referenced and a returned qualified function name
+   *                           cannot otherwise be accessed.
+   * @return the name of the function, qualified by class if it will be inlined to a private,
+   *         nested sub-class
+   */
+  def addNewFunction(
+      funcName: String,
+      funcCode: String,
+      inlineToOuterClass: Boolean = false): String = {
+    // The number of named constants that can exist in the class is limited by the Constant Pool
+    // limit, 65,536. We cannot know how many constants will be inserted for a class, so we use a
+    // threshold of 1600k bytes to determine when a function should be inlined to a private, nested
+    // sub-class.
+    val (className, classInstance) = if (inlineToOuterClass) {
+      outerClassName -> ""
+    } else if (currClassSize > 1600000) {
+      val className = freshName("NestedClass")
+      val classInstance = freshName("nestedClassInstance")
+
+      addClass(className, classInstance)
+
+      className -> classInstance
+    } else {
+      currClass()
+    }
+
+    classSize(className) += funcCode.length
+    classFunctions(className) += funcName -> funcCode
+
+    if (className == outerClassName) {
+      funcName
+    } else {
+
+      s"$classInstance.$funcName"
+    }
+  }
+
+  /**
+   * Instantiates all nested, private sub-classes as objects to the `OuterClass`
+   */
+  private[sql] def initNestedClasses(): String = {
+    // Nested, private sub-classes have no mutable state (though they do reference the outer class'
+    // mutable state), so we declare and initialize them inline to the OuterClass.
+    classes.filter(_._1 != outerClassName).map {
+      case (className, classInstance) =>
+        s"private $className $classInstance = new $className();"
+    }.mkString("\n")
+  }
+
+  /**
+   * Declares all function code that should be inlined to the `OuterClass`.
+   */
+  private[sql] def declareAddedFunctions(): String = {
+    classFunctions(outerClassName).values.mkString("\n")
+  }
+
+  /**
+   * Declares all nested, private sub-classes and the function code that should be inlined to them.
+   */
+  private[sql] def declareNestedClasses(): String = {
+    classFunctions.filterKeys(_ != outerClassName).map {
+      case (className, functions) =>
+        s"""
+           |private class $className {
+           |  ${functions.values.mkString("\n")}
+           |}
+           """.stripMargin
+    }
+  }.mkString("\n")
+
   final val JAVA_BOOLEAN = "boolean"
   final val JAVA_BYTE = "byte"
   final val JAVA_SHORT = "short"
@@ -556,8 +654,7 @@ class CodegenContext {
             return 0;
           }
         """
-      addNewFunction(compareFunc, funcCode)
-      s"this.$compareFunc($c1, $c2)"
+      s"${addNewFunction(compareFunc, funcCode)}($c1, $c2)"
     case schema: StructType =>
       val comparisons = GenerateOrdering.genComparisons(this, schema)
       val compareFunc = freshName("compareStruct")
@@ -573,8 +670,7 @@ class CodegenContext {
             return 0;
           }
         """
-      addNewFunction(compareFunc, funcCode)
-      s"this.$compareFunc($c1, $c2)"
+      s"${addNewFunction(compareFunc, funcCode)}($c1, $c2)"
     case other if other.isInstanceOf[AtomicType] => s"$c1.compare($c2)"
     case udt: UserDefinedType[_] => genComp(udt.sqlType, c1, c2)
     case _ =>
@@ -689,7 +785,6 @@ class CodegenContext {
            |}
          """.stripMargin
         addNewFunction(name, code)
-        name
       }
 
       foldFunctions(functions.map(name => s"$name(${arguments.map(_._2).mkString(", ")})"))
@@ -773,8 +868,6 @@ class CodegenContext {
            |}
            """.stripMargin
 
-      addNewFunction(fnName, fn)
-
       // Add a state and a mapping of the common subexpressions that are associate with this
       // state. Adding this expression to subExprEliminationExprMap means it will call `fn`
       // when it is code generated. This decision should be a cost based one.
@@ -792,7 +885,7 @@ class CodegenContext {
       addMutableState(javaType(expr.dataType), value,
         s"$value = ${defaultValue(expr.dataType)};")
 
-      subexprFunctions += s"$fnName($INPUT_ROW);"
+      subexprFunctions += s"${addNewFunction(fnName, fn)}($INPUT_ROW);"
       val state = SubExprEliminationState(isNull, value)
       e.foreach(subExprEliminationExprs.put(_, state))
     }

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateMutableProjection.scala

@@ -63,21 +63,21 @@ object GenerateMutableProjection extends CodeGenerator[Seq[Expression], MutableP
         if (e.nullable) {
           val isNull = s"isNull_$i"
           val value = s"value_$i"
-          ctx.addMutableState("boolean", isNull, s"this.$isNull = true;")
+          ctx.addMutableState("boolean", isNull, s"$isNull = true;")
           ctx.addMutableState(ctx.javaType(e.dataType), value,
-            s"this.$value = ${ctx.defaultValue(e.dataType)};")
+            s"$value = ${ctx.defaultValue(e.dataType)};")
           s"""
             ${ev.code}
-            this.$isNull = ${ev.isNull};
-            this.$value = ${ev.value};
+            $isNull = ${ev.isNull};
+            $value = ${ev.value};
            """
         } else {
           val value = s"value_$i"
           ctx.addMutableState(ctx.javaType(e.dataType), value,
-            s"this.$value = ${ctx.defaultValue(e.dataType)};")
+            s"$value = ${ctx.defaultValue(e.dataType)};")
           s"""
             ${ev.code}
-            this.$value = ${ev.value};
+            $value = ${ev.value};
            """
         }
     }
@@ -87,7 +87,7 @@ object GenerateMutableProjection extends CodeGenerator[Seq[Expression], MutableP
 
     val updates = validExpr.zip(index).map {
       case (e, i) =>
-        val ev = ExprCode("", s"this.isNull_$i", s"this.value_$i")
+        val ev = ExprCode("", s"isNull_$i", s"value_$i")
         ctx.updateColumn("mutableRow", e.dataType, i, ev, e.nullable)
     }
 
@@ -135,6 +135,9 @@ object GenerateMutableProjection extends CodeGenerator[Seq[Expression], MutableP
           $allUpdates
           return mutableRow;
         }
+
+        ${ctx.initNestedClasses()}
+        ${ctx.declareNestedClasses()}
       }
     """
 

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateOrdering.scala

@@ -179,6 +179,9 @@ object GenerateOrdering extends CodeGenerator[Seq[SortOrder], Ordering[InternalR
           $comparisons
           return 0;
         }
+
+        ${ctx.initNestedClasses()}
+        ${ctx.declareNestedClasses()}
       }"""
 
     val code = CodeFormatter.stripOverlappingComments(

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GeneratePredicate.scala

@@ -72,6 +72,9 @@ object GeneratePredicate extends CodeGenerator[Expression, Predicate] {
           ${eval.code}
           return !${eval.isNull} && ${eval.value};
         }
+
+        ${ctx.initNestedClasses()}
+        ${ctx.declareNestedClasses()}
       }"""
 
     val code = CodeFormatter.stripOverlappingComments(

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateSafeProjection.scala

@@ -49,7 +49,7 @@ object GenerateSafeProjection extends CodeGenerator[Seq[Expression], Projection]
     val output = ctx.freshName("safeRow")
     val values = ctx.freshName("values")
     // These expressions could be split into multiple functions
-    ctx.addMutableState("Object[]", values, s"this.$values = null;")
+    ctx.addMutableState("Object[]", values, s"$values = null;")
 
     val rowClass = classOf[GenericInternalRow].getName
 
@@ -65,10 +65,10 @@ object GenerateSafeProjection extends CodeGenerator[Seq[Expression], Projection]
     val allFields = ctx.splitExpressions(tmp, fieldWriters)
     val code = s"""
       final InternalRow $tmp = $input;
-      this.$values = new Object[${schema.length}];
+      $values = new Object[${schema.length}];
       $allFields
       final InternalRow $output = new $rowClass($values);
-      this.$values = null;
+      $values = null;
     """
 
     ExprCode(code, "false", output)
@@ -184,6 +184,9 @@ object GenerateSafeProjection extends CodeGenerator[Seq[Expression], Projection]
           $allExpressions
           return mutableRow;
         }
+
+        ${ctx.initNestedClasses()}
+        ${ctx.declareNestedClasses()}
       }
     """
 

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

@@ -82,7 +82,7 @@ object GenerateUnsafeProjection extends CodeGenerator[Seq[Expression], UnsafePro
     val rowWriterClass = classOf[UnsafeRowWriter].getName
     val rowWriter = ctx.freshName("rowWriter")
     ctx.addMutableState(rowWriterClass, rowWriter,
-      s"this.$rowWriter = new $rowWriterClass($bufferHolder, ${inputs.length});")
+      s"$rowWriter = new $rowWriterClass($bufferHolder, ${inputs.length});")
 
     val resetWriter = if (isTopLevel) {
       // For top level row writer, it always writes to the beginning of the global buffer holder,
@@ -182,7 +182,7 @@ object GenerateUnsafeProjection extends CodeGenerator[Seq[Expression], UnsafePro
     val arrayWriterClass = classOf[UnsafeArrayWriter].getName
     val arrayWriter = ctx.freshName("arrayWriter")
     ctx.addMutableState(arrayWriterClass, arrayWriter,
-      s"this.$arrayWriter = new $arrayWriterClass();")
+      s"$arrayWriter = new $arrayWriterClass();")
     val numElements = ctx.freshName("numElements")
     val index = ctx.freshName("index")
     val element = ctx.freshName("element")
@@ -321,7 +321,7 @@ object GenerateUnsafeProjection extends CodeGenerator[Seq[Expression], UnsafePro
     val holder = ctx.freshName("holder")
     val holderClass = classOf[BufferHolder].getName
     ctx.addMutableState(holderClass, holder,
-      s"this.$holder = new $holderClass($result, ${numVarLenFields * 32});")
+      s"$holder = new $holderClass($result, ${numVarLenFields * 32});")
 
     val resetBufferHolder = if (numVarLenFields == 0) {
       ""
@@ -402,6 +402,9 @@ object GenerateUnsafeProjection extends CodeGenerator[Seq[Expression], UnsafePro
           ${eval.code.trim}
           return ${eval.value};
         }
+
+        ${ctx.initNestedClasses()}
+        ${ctx.declareNestedClasses()}
       }
       """
 

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

@@ -93,7 +93,7 @@ private [sql] object GenArrayData {
     if (!ctx.isPrimitiveType(elementType)) {
       val genericArrayClass = classOf[GenericArrayData].getName
       ctx.addMutableState("Object[]", arrayName,
-        s"this.$arrayName = new Object[${numElements}];")
+        s"$arrayName = new Object[${numElements}];")
 
       val assignments = elementsCode.zipWithIndex.map { case (eval, i) =>
         val isNullAssignment = if (!isMapKey) {
@@ -340,7 +340,7 @@ case class CreateNamedStruct(children: Seq[Expression]) extends CreateNamedStruc
   override def doGenCode(ctx: CodegenContext, ev: ExprCode): ExprCode = {
     val rowClass = classOf[GenericInternalRow].getName
     val values = ctx.freshName("values")
-    ctx.addMutableState("Object[]", values, s"this.$values = null;")
+    ctx.addMutableState("Object[]", values, s"$values = null;")
 
     ev.copy(code = s"""
       $values = new Object[${valExprs.size}];""" +
@@ -357,7 +357,7 @@ case class CreateNamedStruct(children: Seq[Expression]) extends CreateNamedStruc
         }) +
       s"""
         final InternalRow ${ev.value} = new $rowClass($values);
-        this.$values = null;
+        $values = null;
       """, isNull = "false")
   }