mgaido91 commented on a change in pull request #21599: [SPARK-26218][SQL]
Overflow on arithmetic operations returns incorrect result
URL: https://github.com/apache/spark/pull/21599#discussion_r299659190
##########
File path:
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/arithmetic.scala
##########
@@ -129,17 +131,41 @@ abstract class BinaryArithmetic extends BinaryOperator
with NullIntolerant {
def calendarIntervalMethod: String =
sys.error("BinaryArithmetics must override either calendarIntervalMethod
or genCode")
+ def checkOverflowCode(result: String, op1: String, op2: String): String =
+ sys.error("BinaryArithmetics must override either checkOverflowCode or
genCode")
+
override def doGenCode(ctx: CodegenContext, ev: ExprCode): ExprCode =
dataType match {
case _: DecimalType =>
defineCodeGen(ctx, ev, (eval1, eval2) =>
s"$eval1.$decimalMethod($eval2)")
case CalendarIntervalType =>
defineCodeGen(ctx, ev, (eval1, eval2) =>
s"$eval1.$calendarIntervalMethod($eval2)")
+ // In the following cases, overflow can happen, so we need to check the
result is valid.
+ // Otherwise we throw an ArithmeticException
// byte and short are casted into int when add, minus, times or divide
case ByteType | ShortType =>
- defineCodeGen(ctx, ev,
- (eval1, eval2) => s"(${CodeGenerator.javaType(dataType)})($eval1
$symbol $eval2)")
+ nullSafeCodeGen(ctx, ev, (eval1, eval2) => {
+ val overflowCheck = if (checkOverflow) {
+ checkOverflowCode(ev.value, eval1, eval2)
+ } else {
+ ""
+ }
+ s"""
+ |${ev.value} = (${CodeGenerator.javaType(dataType)})($eval1 $symbol
$eval2);
+ |$overflowCheck
+ """.stripMargin
+ })
case _ =>
- defineCodeGen(ctx, ev, (eval1, eval2) => s"$eval1 $symbol $eval2")
+ nullSafeCodeGen(ctx, ev, (eval1, eval2) => {
+ val overflowCheck = if (checkOverflow) {
+ checkOverflowCode(ev.value, eval1, eval2)
+ } else {
+ ""
+ }
+ s"""
+ |${ev.value} = $eval1 $symbol $eval2;
Review comment:
yes, this is good point actually. As I mentioned in
https://github.com/apache/spark/pull/21599#issuecomment-505924180, the issue
with this is that `addExact` exists only for few datatypes. So we should
specialize many cases in the code here. So here we have a tradeoff between
performances (in the check overflow case only) and code maintainability. What
is your opinion on this?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
