[
https://issues.apache.org/jira/browse/HIVE-12315?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14986627#comment-14986627
]
Gopal V commented on HIVE-12315:
--------------------------------
[~mmccline]: the loop-hoist was correct, please review attached patch.
The divbyZero codepath unconditionally resets .noNulls=false, without
considering that the .isNull[] state is not reset when the input had .noNulls =
true.
> vectorization_short_regress.q has a wrong result issue for a double
> calculation
> -------------------------------------------------------------------------------
>
> Key: HIVE-12315
> URL: https://issues.apache.org/jira/browse/HIVE-12315
> Project: Hive
> Issue Type: Bug
> Components: Vectorization
> Reporter: Matt McCline
> Assignee: Gopal V
> Priority: Critical
> Attachments: HIVE-12315.1.patch, vectorization_short_regress_bug.q
>
>
> I suspect it is related to the fancy optimizations in vectorized double
> divide that try to quickly process the batch without checking each row for
> null.
> {code}
> public static void setNullAndDivBy0DataEntriesDouble(
> DoubleColumnVector v, boolean selectedInUse, int[] sel, int n,
> DoubleColumnVector denoms) {
> assert v.isRepeating || !denoms.isRepeating;
> v.noNulls = false;
> double[] vector = denoms.vector;
> if (v.isRepeating && (v.isNull[0] = (v.isNull[0] || vector[0] == 0))) {
> v.vector[0] = DoubleColumnVector.NULL_VALUE;
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
