David Mollitor created SPARK-59379:
--------------------------------------
Summary: Add ASCII fast-path to UTF8String character/byte position
lookups
Key: SPARK-59379
URL: https://issues.apache.org/jira/browse/SPARK-59379
Project: Spark
Issue Type: Improvement
Components: Spark Core
Affects Versions: 4.1.0
Reporter: David Mollitor
h2. Summary
Several {{UTF8String}} methods convert a character index to/from a byte offset
by walking the string one code point at a time ({{{}i +=
numBytesForFirstByte(getByte(i){}}}). For a full-ASCII string every code point
is exactly one byte, so *character index == byte index* and that scan can be
replaced with O(1) arithmetic.
This adds an ASCII fast-path to the four char<->byte "locate" methods:
{{{}substring(int, int){}}}, {{{}getChar(int){}}}, {{{}charPosToByte(int){}}},
and {{{}bytePosToChar(int){}}}.
h2. Details
The fast-path is gated on the cached ASCII flag being {*}already warm{*}, read
without forcing a scan:
{code:java}
if (isFullAscii == IsFullAscii.FULL_ASCII) {
// ASCII: char index == byte index, so locate directly.
...
}
{code}
* This is a pure win with {*}no regression{*}: when the flag is cold
({{{}UNKNOWN{}}}) the methods fall back to the existing per-code-point scan. It
deliberately does NOT call {{{}isFullAscii(){}}}, which would force a full scan
and could regress cold long strings (e.g. {{{}SUBSTR(s, 1, 5){}}}).
* Arithmetic replicates the scan's behavior exactly: a negative index clamps
to 0, the {{until == Integer.MAX_VALUE}} sentinel maps to {{{}numBytes{}}}, and
endpoints clamp to {{{}numBytes{}}}. For example {{substring}} computes {{j =
max(start, 0)}} and {{{}i = (until == Integer.MAX_VALUE) ? numBytes :
min(until, numBytes){}}}; {{getChar}} returns {{codePointFrom(charIndex)}}
directly (the existing {{Objects.checkIndex}} guarantees the index is in range).
* Non-ASCII and invalid-UTF-8 strings have the flag set to {{NOT_ASCII}} when
warmed, so they take the unchanged scan path.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
