AdamGS commented on code in PR #9625:
URL: https://github.com/apache/arrow-rs/pull/9625#discussion_r3347406250
##########
arrow-select/src/take.rs:
##########
@@ -495,99 +495,126 @@ fn take_bytes<T: ByteArrayType, IndexType:
ArrowPrimitiveType>(
array: &GenericByteArray<T>,
indices: &PrimitiveArray<IndexType>,
) -> Result<GenericByteArray<T>, ArrowError> {
+ let mut values: Vec<u8> = Vec::new();
let mut offsets = Vec::with_capacity(indices.len() + 1);
offsets.push(T::Offset::default());
let input_offsets = array.value_offsets();
let mut capacity = 0;
let nulls = take_nulls(array.nulls(), indices);
- let (offsets, values) = if array.null_count() == 0 && indices.null_count()
== 0 {
- offsets.reserve(indices.len());
- for index in indices.values() {
- let index = index.as_usize();
- capacity += input_offsets[index + 1].as_usize() -
input_offsets[index].as_usize();
- offsets.push(
- T::Offset::from_usize(capacity)
- .ok_or_else(|| ArrowError::OffsetOverflowError(capacity))?,
- );
- }
- let mut values = Vec::with_capacity(capacity);
-
- for index in indices.values() {
- values.extend_from_slice(array.value(index.as_usize()).as_ref());
- }
- (offsets, values)
- } else if indices.null_count() == 0 {
- offsets.reserve(indices.len());
- for index in indices.values() {
- let index = index.as_usize();
- if array.is_valid(index) {
- capacity += input_offsets[index + 1].as_usize() -
input_offsets[index].as_usize();
+ // Branch on output nulls — `None` means every output slot is valid.
+ match nulls.as_ref().filter(|n| n.null_count() > 0) {
+ // Fast path: no nulls in output, every index is valid.
+ None => {
+ for index in indices.values() {
+ let index = index.as_usize();
+ let start = input_offsets[index].as_usize();
+ let end = input_offsets[index + 1].as_usize();
+ capacity += end - start;
+ offsets.push(
+ T::Offset::from_usize(capacity)
+ .ok_or_else(||
ArrowError::OffsetOverflowError(capacity))?,
+ );
}
- offsets.push(
- T::Offset::from_usize(capacity)
- .ok_or_else(|| ArrowError::OffsetOverflowError(capacity))?,
- );
- }
- let mut values = Vec::with_capacity(capacity);
- for index in indices.values() {
- let index = index.as_usize();
- if array.is_valid(index) {
- values.extend_from_slice(array.value(index).as_ref());
- }
- }
- (offsets, values)
- } else if array.null_count() == 0 {
- offsets.reserve(indices.len());
- for (i, index) in indices.values().iter().enumerate() {
- let index = index.as_usize();
- if indices.is_valid(i) {
- capacity += input_offsets[index + 1].as_usize() -
input_offsets[index].as_usize();
+ values.reserve(capacity);
+
+ let dst = values.spare_capacity_mut();
+ debug_assert!(dst.len() >= capacity);
+ let mut offset = 0;
+
+ for index in indices.values() {
+ // SAFETY: in-bounds proven by the first loop's bounds-checked
offset access.
+ // dst asserted above to include the required capacity.
+ unsafe {
+ let data: &[u8] =
array.value_unchecked(index.as_usize()).as_ref();
+ std::ptr::copy_nonoverlapping(
+ data.as_ptr(),
+ dst[offset..].as_mut_ptr().cast::<u8>(),
+ data.len(),
+ );
+ offset += data.len();
+ }
}
- offsets.push(
- T::Offset::from_usize(capacity)
- .ok_or_else(|| ArrowError::OffsetOverflowError(capacity))?,
- );
- }
- let mut values = Vec::with_capacity(capacity);
- for (i, index) in indices.values().iter().enumerate() {
- if indices.is_valid(i) {
-
values.extend_from_slice(array.value(index.as_usize()).as_ref());
+ // SAFETY: wrote exactly `capacity` bytes above; reserved on line
above.
+ unsafe {
+ values.set_len(capacity);
}
}
- (offsets, values)
- } else {
- let nulls = nulls.as_ref().unwrap();
- offsets.reserve(indices.len());
- for (i, index) in indices.values().iter().enumerate() {
- let index = index.as_usize();
- if nulls.is_valid(i) {
- capacity += input_offsets[index + 1].as_usize() -
input_offsets[index].as_usize();
+ // Nullable path: only process valid (non-null) output positions.
+ Some(output_nulls) => {
+ let mut ranges = Vec::with_capacity(indices.len() -
output_nulls.null_count());
Review Comment:
I think I've tried to avoid it, but I couldn't find anything I liked or
performed better.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
