[SPARK-29100][SQL] Fix compilation error in codegen with switch from InSet expression

[viirya]

What changes were proposed in this pull request?

When InSet generates Java switch-based code, if the input set is empty, we don't generate switch condition, but a simple expression that is default case of original switch.

Why are the changes needed?

SPARK-26205 adds an optimization to InSet that generates Java switch condition for certain cases. When the given set is empty, it is possibly that codegen causes compilation error:

[info] - SPARK-29100: InSet with empty input set *** FAILED *** (58 milliseconds)                                      
[info]   Code generation of input[0, int, true] INSET () failed:                                                                        
[info]   org.codehaus.janino.InternalCompilerException: failed to compile: org.codehaus.janino.InternalCompilerException: Compiling "GeneratedClass" in "generated.java": Compiling "apply(java.lang.Object _i)"; apply(java.lang.Object _i): Operand stack inconsistent at offset 45: Previous size 0, now 1                                                                                                           
[info]   org.codehaus.janino.InternalCompilerException: failed to compile: org.codehaus.janino.InternalCompilerException: Compiling "GeneratedClass" in "generated.java": Compiling "apply(java.lang.Object _i)"; apply(java.lang.Object _i): Operand stack inconsistent at offset 45: Previous size 0, now 1                                                                                                           
[info]         at org.apache.spark.sql.catalyst.expressions.codegen.CodeGenerator$.org$apache$spark$sql$catalyst$expressions$codegen$CodeGenerator$$doCompile(CodeGenerator.scala:1308)                                                                                        
[info]         at org.apache.spark.sql.catalyst.expressions.codegen.CodeGenerator$$anon$1.load(CodeGenerator.scala:1386)               
[info]         at org.apache.spark.sql.catalyst.expressions.codegen.CodeGenerator$$anon$1.load(CodeGenerator.scala:1383)

Does this PR introduce any user-facing change?

Yes. Previously, when users have InSet against an empty set, generated code causes compilation error. This patch fixed it.

How was this patch tested?

Unit test added.

[viirya]

cc @cloud-fan @dongjoon-hyun @maropu

[dongjoon-hyun]

Thank you for fixing this, @viirya .

[SparkQA]

Test build #110660 has finished for PR 25806 at commit 5ad7538.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[maropu]

We don't need to compute caseBranches, too, if hset is empty?

    val valueGen = child.genCode(ctx)
    val switchCode = if (hset.isEmpty) {
      val caseValuesGen = hset.filter(_ != null).map(Literal(_).genCode(ctx))
      val caseBranches = caseValuesGen.map(literal =>
        code"""
          case ${literal.value}:
            ${ev.value} = true;
            break;
       """)

      code"""
        switch (${valueGen.value}) {
          ${caseBranches.mkString("\n")}
          default:
            ${ev.isNull} = $hasNull;
        }
       """
    } else {
      s"${ev.isNull} = $hasNull;"
    }

[maropu]

btw, can we fold this expr in the optimizer when hset is empty?

[cloud-fan]

if hset is empty, the computation of caseBranches is noop, right?

[cloud-fan]

This is a special case. Usually InSet wouldn't have empty set, as we have a threshold to convert In to Inset.

[maropu]

Ur, I see.

[maropu]

LGTM except for one minor comment.

[cloud-fan]

thanks, merging to master!

[viirya]

thanks @maropu @cloud-fan @dongjoon-hyun