[
https://issues.apache.org/jira/browse/SPARK-59390?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Uroš Bojanić reassigned SPARK-59390:
------------------------------------
Assignee: David Mollitor
> Build padding in UTF8String.lpad/rpad with a shared repeat-fill helper
> ----------------------------------------------------------------------
>
> Key: SPARK-59390
> URL: https://issues.apache.org/jira/browse/SPARK-59390
> Project: Spark
> Issue Type: Improvement
> Components: Spark Core
> Affects Versions: 4.1.0
> Reporter: David Mollitor
> Assignee: David Mollitor
> Priority: Minor
> Labels: pull-request-available
>
> h2. Summary
> {{lpad}} and {{rpad}} build the padding by copying the pad string {{count}}
> times in a loop -- one {{copyMemory}} per repetition, i.e. O(count) copy
> calls. {{repeat()}} already solves the same "repeat these bytes N times"
> problem more efficiently, with a single-byte {{Arrays.fill}} fast path and
> exponential doubling. This factors that logic into a shared private helper
> ({{fillRepeated}}) used by {{repeat()}}, {{lpad}} and {{rpad}}.
> h2. Details
> * Single-byte pattern -- the common padding case (space, {{0}}, {{?}}) --
> fills the region with a single {{Arrays.fill}}.
> * Multi-byte pattern -- seed one copy, then exponential doubling: O(log
> count) copies instead of O(count).
> * The total number of bytes copied is unchanged; only the number of copy
> calls drops. {{repeat()}} now delegates to the shared helper, and its
> previous inline doubling and single-byte branch are removed, leaving a single
> implementation.
> * The benefit shows up when padding a short value to a large width with a
> short pad (large {{count}}); ordinary small-width padding is unaffected.
> This is a modest, targeted change.
> h2. Does this PR introduce _any_ user-facing change?
> No. Behavior is identical; only the way the padding bytes are written into
> the result buffer changes.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
