[
https://issues.apache.org/jira/browse/SPARK-59969?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
ASF GitHub Bot updated SPARK-59969:
-----------------------------------
Labels: pull-request-available (was: )
> Skip the redundant bounds check on internal UTF8String byte reads
> -----------------------------------------------------------------
>
> Key: SPARK-59969
> URL: https://issues.apache.org/jira/browse/SPARK-59969
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 4.2.0
> Reporter: David Mollitor
> Priority: Minor
> Labels: pull-request-available
>
> {{UTF8String.getByte(int)}} checks that the index is in {{[0, numBytes)}} and
> returns 0 otherwise. {{UTF8String}} also calls it internally from the loops
> behind {{numChars}}, {{isValid}} / {{makeValid}}, {{isFullAscii}},
> {{substring}}, {{charPosToByte}} / {{bytePosToChar}}, {{indexOf}},
> {{reverse}}, the {{trim}} family, {{toInt}} / {{toLong}}, ASCII case
> conversion, {{findInSet}}, {{soundex}} and the code point iterators. In those
> loops the loop condition already keeps the index in range, so the check is
> redundant. The JIT can't always remove it, especially in character walks,
> where each step advances by the width of the current character.
> This adds a private {{getByteUnchecked(int)}}, which asserts that the index
> is in range (Spark tests run with {{-ea}}), and makes {{getByte}} call it
> after its own check. The 45 internal call sites whose loop provably keeps the
> index in range switch to {{getByteUnchecked}}. The other 11 stay on
> {{getByte}} because they can index past the end, or depend on an invariant
> outside the loop:
> * {{contains}}, which SPARK-59967 replaces
> * {{getChar}}
> * {{trimTrailingSpaces}}
> * {{prevCharStart}}
> * one read in {{indexOf(pattern, start, occurrence)}} that can run past the
> end on truncated UTF-8
> * the walks in {{levenshteinDistance}}
> The behavior of the public {{getByte}} does not change.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
