[SPARK-35329] Split generated switch code into pieces in ExpandExec

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 3.2.0
• Fix Version/s: 3.2.0
• Component/s: SQL
• Assignee: Takeshi Yamamuro
• Reporter: Takeshi Yamamuro
• Created: 06/May/21 15:12
• Updated: 14/May/21 00:54
• Resolved: 14/May/21 00:54

Description

This ticket aim at splitting generated switch code into smaller ones in `ExpandExec`. In the current master, even a simple query like the one below generates a large method whose size (`maxMethodCodeSize:7448`) is close to `8000` (`CodeGenerator.DEFAULT_JVM_HUGE_METHOD_LIMIT`);

scala> val df = Seq(("2016-03-27 19:39:34", 1, "a"), ("2016-03-27 19:39:56", 2, "a"), ("2016-03-27 19:39:27", 4, "b")).toDF("time", "value", "id")
scala> val rdf = df.select(window($"time", "10 seconds", "3 seconds", "0 second"), $"value").orderBy($"window.start".asc, $"value".desc).select("value")
scala> sql("SET spark.sql.adaptive.enabled=false")
scala> import org.apache.spark.sql.execution.debug._ 
scala> rdf.debugCodegen

Found 2 WholeStageCodegen subtrees.
== Subtree 1 / 2 (maxMethodCodeSize:7448; maxConstantPoolSize:189(0.29% used); numInnerClasses:0) ==
                                    ^^^^
*(1) Project [window#34.start AS _gen_alias_39#39, value#11]
+- *(1) Filter ((isnotnull(window#34) AND (cast(time#10 as timestamp) >= window#34.start)) AND (cast(time#10 as timestamp) < window#34.end))
   +- *(1) Expand [List(named_struct(start, precisetimestampcon...

Issue Links

• links to: [Github] Pull Request #32457 (maropu)