[
https://issues.apache.org/jira/browse/IMPALA-7923?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17154004#comment-17154004
]
ASF subversion and git services commented on IMPALA-7923:
---------------------------------------------------------
Commit 45c105d71d47c4c57e042b9bf8a0d8d8044083bc in impala's branch
refs/heads/master from Daniel Becker
[ https://gitbox.apache.org/repos/asf?p=impala.git;h=45c105d ]
IMPALA-7923: DecimalValue should be marked as packed
IMPALA-7473 and IMPALA-9111 were symptoms of a more general problem that
DecimalValue is not guaranteed to be aligned by the Impala runtime but
the compiler assumes it is and under some circumstances, it will emit
code for aligned loads to value_ when value_ is an int128.
This commit marks DecimalValue as packed so that the compiler does not
assume any alignment.
TODO: Maybe benchmark if this introduces performance regressions, but it
shouldn't.
Change-Id: I55f936a4f4f4b5faf129a9265222e64fc486b8ed
Reviewed-on: http://gerrit.cloudera.org:8080/16134
Reviewed-by: Impala Public Jenkins <[email protected]>
Tested-by: Impala Public Jenkins <[email protected]>
> DecimalValue should be marked as packed
> ---------------------------------------
>
> Key: IMPALA-7923
> URL: https://issues.apache.org/jira/browse/IMPALA-7923
> Project: IMPALA
> Issue Type: Improvement
> Components: Backend
> Affects Versions: Impala 3.1.0
> Reporter: Tim Armstrong
> Assignee: Daniel Becker
> Priority: Major
>
> IMPALA-7473 was a symptom of a more general problem that DecimalValue is not
> guaranteed to be aligned by the Impala runtime, but the class is not marked
> as packed and, under some circumstances, GCC will emit code for aligned loads
> to value_ when value_ is an int128.
> Testing helps confirm that the compiler does not emit the problematic loads
> in practice, but it would be better to mark the struct as packed.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
